acceptodds
Under review as a conference paper at ICLR 2027

Accumulation Is Maintenance: Anti-Preferential Placement and Online Structural Curation for Hierarchical Agent Memory

Abstract

Long-term memory systems for LLM agents have converged on online insertion: each new observation is routed to a location in a hierarchy or graph, and the structure is then left alone; at best, a single top layer is periodically repartitioned. We show that this neglect is not benign. On a multi-topic accumulation workload derived from real usage, undisciplined insertion produces structural pathologies: a single early node absorbs 64% of all records through preferential attachment, topics fragment across near-duplicate siblings, and the hierarchy stays flat. Recall-centric benchmarks such as LoCoMo and LongMemEval miss these pathologies, because recall holds as long as an answer is somewhere in memory. We present Gardener, a hierarchical memory system built on the principle that accumulation is maintenance, not placement. Gardener combines anti-preferential placement, which deterministically removes the volume advantage behind rich-get-richer growth; an online curation loop that merges, splits, and relocates nodes under strict guards (original records are never modified, and every write is executed by deterministic code); and a three-way ingestion triage that moves abstention from recall time to accumulation time. We also propose an accumulation-quality evaluation protocol pairing structural-health metrics with writer attribution (the oracle-to-e2e gap). In same-condition comparisons on an English-translated workload against online tree, graph, and extraction baselines (MemTree, A-MEM, Mem0), Gardener matches embedding-based recall while its non-destructive supersede updates eliminate stale-value citations on knowledge-update probes: 0% in every run, against roughly 50% for every baseline. Because the guards are deterministic code, their bound survives a backbone change that model behavior does not: unguarded accumulation ranges from the 64% collapse on the original backbone to moderate drift on a newer one, while guarded accumulation stays within the health gate in both regimes, curation repairing what the newer backbone's intake degrades.

Primary area:
foundation or frontier models, including LLMs
Keywords:
agent memory, long-term memory, hierarchical memory, LLM agents, memory maintenance, online structural curation, evaluation protocol

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DuiZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that online hierarchical memory systems for LLM agents decide where each new record goes but never revisit that structure afterward. It claims this lets structural problems build up: a "magnet" node that pulls in most records through preferential attachment, near-duplicate sibling branches (fragmentation), and flat trees. Recall benchmarks such as LoCoMo and LongMemEval don't detect these problems. The proposed system, Gardener, has four parts:

  1. Anti-preferential placement. Candidate nodes are scored by code using per-keyword breadth voting divided by √(|v|+1). The LLM may only choose among these candidates.
  2. An online curation loop. It relocates, clusters, merges, splits, renames and cross-links nodes on nightly and reactive triggers. The LLM proposes each change, and deterministic code validates and applies it. Original records are never changed.
  3. Three-way ingestion triage. Each record is either placed, sent back to the user as a question ("ask-back"), or held as unclassified.
  4. Non-destructive "supersede" updates. Stale records are marked invalid, in the style of Zep/Graphiti bi-temporal invalidation, rather than rewritten or deleted.

The paper also proposes a set of structural-health metrics: magnet occupancy M, a title-Jaccard fragmentation index F, depth distribution, and unclassified rate. It pairs them with an oracle-vs-e2e "writer attribution" protocol taken from RoboMME.

Validation. The main tests use a 31-message multi-topic corpus derived from real usage. The corpus was originally in Korean, was translated to English, and is withheld. There are also 8 self-authored fixtures (51 questions). Baselines are MemTree (a reimplementation), Mem0 and A-MEM.

Results. Gardener's e2e recall (84.3%) is on par with the baselines (80–82%) and identical to the authors' own flat, no-tree control. Gardener cites stale values 0% of the time, against about 50% for the baselines. Its magnet occupancy is 31% at intake and 22% after one curation pass, against 42% for MemTree. The headline 64% collapse comes from the original backbone (Gemini 2.5 Flash), which was retired during the study. On the current backbone, turning off all guards gives only 33%. Out-of-domain probes are also reported: full LoCoMo with token-F1 0.331, and the LongMemEval knowledge-update and abstention subsets with 52/102.

Critical strengths
  • S1. A real and under-measured problem, with a reasonable deterministic yardstick. Recall-only benchmarks can't see structural degradation, because an answer anywhere in memory still counts. The MemTree sweep (Fig. 4) gives concrete support: depth-adaptive thresholds alone trade one problem for another, either a magnet (θ₀=0.3 → M=55%) or deep fragmented chains (λ=0.1 → 24 branches, depth 8). The structural metrics are computed without an LLM, so they can be reproduced across systems. This framing could be useful for the memory-systems community even if the specific system turns out not to be.

S2. Non-destructive, code-applied supersede updates clearly remove stale-value citations. This is the paper's strongest empirical result: 0/18 stale citations in every run, against 50–56% for MemTree, Mem0 and A-MEM, with old values still answerable through a history query. To their credit, the authors attribute this to the supersede guard rather than to the tree, since the flat control also gets 0%. The general design rule ("the LLM proposes, code executes") is sound and practical. Some caveats apply to the size of this gain:

  • MemTree has no update mechanism at all, so part of the gap is built into the comparison.
  • Gardener declines to answer 15.7% of the time, while the baselines never do, and a non-answer can never cite a stale value.
  • The probe has 18 author-written questions in which the updates are stated explicitly.
  • On the real LongMemEval knowledge-update subset (38/72), stale-value citations are not reported.
Critical weaknesses

W1. On the primary backbone, the guards are not shown to bound structure, and the paper's own numbers contradict its gate claims.

  • The intake numbers fail the gates. The gates are M ≤ 0.30 and unclassified ≤ 0.10 (§3.3). Intake reaches M = 31% (27–55% over 3 runs) and unfiled 33%, failing both. Yet three places claim otherwise:
  • §6.2 says intake "stays within the health gate (magnet 31% at intake…)";
  • §6.5 says all backbones "meet the structural-health gates";
  • App. E lists "magnet 31%, unfiled 33%" and then says generation shifts surface "not as gate violations."

Only after curation (22%, 4%) are both gates met. The abstract should state that it refers to the post-curation state.

  • Removing the guards makes no measurable difference. With all six guards off, M is 33%, against 31% with them on (Fig. 2). Removing anti-preferential scoring, contribution (1), leaves M unchanged on both backbones (31→31 and 20→20). Single ablations range from 25% to 42% at n=1, which is within run-to-run spread.
  • The motivating collapse may not exist on current models. The 64% collapse appears only on a retired backbone, and only with all guards removed together. No bound is ever stated or measured: √-normalization weakens the advantage of large nodes but does not cap occupancy. The gates were also calibrated on the authors' own system, which makes passing them partly circular.

The abstract presents the backbone change as support for deterministic guards. In fact it weakens the intake-side contributions, (1) anti-preferential placement and (3) triage: neither changes intake M measurably on the current backbone (removing scoring gives 31→31; removing ask-back gives 25 vs 31). Only contribution (2), curation, has support, and only on one 31-record corpus (M 31% → 22%, unfiled 33% → 4%).

W2. Nothing shows that structural health matters for any task, and on the primary suite the tree adds nothing measurable over a flat store. The paper says poor structure hurts "branch-level context injection, summary coherence, and navigability," but none of these is measured. On the primary suite, Gardener e2e and the unstructured flat control are identical on every reported metric: recall 84.3%, stale citations 0%, avoidance 15.7%. The freshness benefit comes from supersede, which is independent of the hierarchy. So the experiments don't show that lower M or F leads to better agent behavior, and the structural metrics stay proxies with no demonstrated validity.

Two further problems weaken the metrics themselves:

  • F = 0 for every tree system in Table 2, so the fragmentation index doesn't distinguish between systems.
  • M can be trivially improved, for example by deep chains (MemTree λ=0.1 reaches 19%, below Gardener's intake).

A convincing version would vary structural health under controlled conditions and show a downstream effect, such as summary faithfulness, navigation or tool-use success, or retrieval cost.

Questions for authors

Clarifications. These can be answered from existing results, and the answers bear directly on W1 and S2:

  1. Is the health gate meant to apply only after curation? If so, the claims in §6.2, §6.5 and App. E should say so, and the intake-side guards would then have no measured effect. If not, how is intake M = 31% (up to 55%) with 33% unfiled "within the gate"? (This affects W1.)
  2. Is the curated 22% / 4% a median over independent runs or a single manual pass (Fig. 3)? Curation is the only component with measured support, so its variance matters. (If this is a single run, W1 and the rating are reinforced.)
  3. On the 18 knowledge-update questions, how many did Gardener decline to answer in each run, and what are its stale-value and accuracy rates on the questions it did answer? (If many were declined, the strength in S2 shrinks.)
  4. The Ethics Statement says no real-user data was used, yet the corpus is "derived from real usage." Which is correct? (This affects the ethics assessment.)

Additional evidence that could change my rating:

  1. With at least 5 runs per ablation cell and intervals, does any intake guard, in particular anti-preferential scoring, change M distinguishably from noise on the current backbone? Can the "bound enforced in code" be stated formally? (A positive answer would address W1.)
  2. Is there any downstream measure, such as summary faithfulness, navigation success, or token or latency cost, on which the tree beats the flat store? (A positive answer would address W2.)
5 comments
AuthorsOpenReview ↗

Reply to DuiZ (1 of 2): corrections to the gate claims; E1 answered at n = 5, the hierarchy nudge, not scoring, prevents collapse, and the code bound is a score bound

Note: See our corrections of 15 Nov below.

We agree with W1 on the gate claims and on the single-run evidence. On the collapse itself, the public rerun below shows that it exists on the current backbone: with one guard off, a 586-turn conversation collapses to 7 nodes in five of five runs. We set the private corpus aside and rebuilt the evidence on public benchmarks during the discussion period, as E1 and E2 ask. On LoCoMo conversation 2 (586 turns) we ran seven intake cells at five runs each with ranges and an exact permutation test, replicated the decisive cell on two further conversations, curated five stores over three passes, and measured branch-level injection on four store shapes. On the LongMemEval knowledge-update subset (72 questions) we ran Gardener, a flat control and a single-call reader over the Gardener store alongside MemTree, Mem0 and A-MEM, all under one grader, with stale citations post-graded among answered questions. Every number below comes from these runs. Terms: magnet share is the fraction of records in the largest node; unclassified records are those left at the root; fragment pairs are sibling nodes with near-duplicate titles; an ask-back is a placement question put to the user. The hierarchy nudge asks the writer to open a sub-branch once a branch holds more than 8 direct records, and title discipline forbids sentence-like branch titles.

Q1, W1-a, W1-d. Five passages (abstract, Sections 6.2 and 6.5, conclusion, Appendix E) are wrong: they say that guarded intake meets the health gate, but only the post-curation state does. At intake, magnet share is 31% and unclassified share 33%; after curation they are 22% and 4%. We also withdraw the claim that the guards are model-independent, since the writer's recording behavior differs sharply across backbones. The thresholds are diagnostics, not controls: they were fixed in July 2026 before the evaluation, no code reads them, and the stores built from 10 of 10 LoCoMo dialogues and 100 of 102 LongMemEval questions meet the same thresholds.

W1-b/c/e, E1. On the evidence below, the guard claim narrows to one mechanism, the hierarchy nudge. First, the "bound enforced in code" is a bound on a candidate's score, not on a node's share of records. The keyword vote total raw(v) of candidate node v is at most 42, so score(v) is at most 42/sqrt(|v|+1) plus a recency term, where |v| is the node's record count. We reword the text to "volume-blind, size-penalized scoring". Second, sibling merge and bloat split (two of the six guards) run only during curation, so ablating them during ingestion could not show anything. On public data, two cells collapse the tree, one spreads it, and the other two match the baseline:

cell (n = 5)magnet median [range]p vs baselinenodes
baseline0.040 [0.032–0.050]83
all four intake guards off1.00 [0.73–1.00]0.0486
hierarchy nudge off0.53 [0.47–0.54]0.0487
scoring term (anti-preferential) off0.042 [0.034–0.059]1.086
ask-back off0.040 [0.030–0.049]1.085
title discipline off0.031 [0.025–0.038]0.04899

p = 0.048 is the smallest value an exact permutation test can give with five runs per cell. Turning off the hierarchy nudge alone reproduces the collapse. Three further runs each on conversations 3 and 6 replicate it under a rule fixed before the runs: magnet share 0.48 versus 0.044 on conversation 3 and 0.96 versus 0.034 on conversation 6, with every nudge-off run above every baseline run. Title discipline off spreads the tree instead (99 nodes, lower magnet share). The nudge is the one guard that reacts to branch size. We attribute the effect to this size feedback, not to the scoring term, and restrict the claim accordingly. The ask-back guard, contribution (3), leaves magnet share unchanged (p = 1.0) and acts on the unclassified share instead (Q2).

Q2. The curated 22% and 4% are one manual pass, not a median; Table 2 and Figure 3 now say so, and the node-count entry (13 to 11) pairs the three-run intake median with that single pass. On public stores curation does merge and split, which the submitted run never did (it only reassigned 7 records). On five public stores with three passes each, pass one reassigns a median of 10 records [4–16], merges 15 nodes [12–23], splits 7 nodes [6–8], and creates 2 new branches. Fragment pairs fall from 5 to 1, unclassified share from 2.1% to 0.2%, and node count from 83 to 69. Magnet share rises from 0.040 to 0.059 (0.088 after pass three) because merges enlarge the largest node; it stays under the 0.30 gate. Splitting is capped at two nodes per pass, so three passes do not converge. The curation claim now rests on the fragment-pair reduction. The drop in unclassified share comes mostly from resolved ask-backs, not from curation: turning ask-backs off alone gives 0.2% (p = 0.048).

(Continued in part 2.)

AuthorsOpenReview ↗

Reply to DuiZ (2 of 2): curation is one pass, now measured over five stores; E2: the tree beats flat on LongMemEval, and structure sets injection recall; stale rates; ethics resolved

Note: See our corrections of 15 Nov below.

(Part 2 of 2; terms and conditions are in part 1.)

W2, E2. Two downstream measures now separate the tree from the flat store. On LongMemEval, with the same writer and reader, the tree store answers 63 of 72 knowledge-update questions and the flat store 41, and the stale rates are 6.9% versus 28.3% (Q3 below). Under branch-level injection the session tree delivers the evidence. The submitted parity arose because the reader never used the tree: it searches all records, and branch injection was off. When context is injected by branch (each node contributes its roll-up, a digest capped at 20 lines), the structure decides which records reach the model. We stored the same 586 records in four shapes: a session tree (one node per session), one magnet node (every record in one node), a deep chain (each session nested under the previous one) and a flat store (no nodes). For each question we pinned (injected) the branch that holds the evidence turn:

variantevidence turn inside the injected block (193 questions, deterministic)
session tree67.9%
magnet0.5%
deep chain1.0%
flat, no pinn/a

Correction (13 Nov). This table first included pinned-reader F1 and fallback counts. They are withdrawn: in those runs, between 24 and 33 of the 56 reader replies per run were error replies returned while the model API was rate-limited, and our harness graded them as answers. The deterministic column above is unaffected. These runs are being repeated with retries, and corrected values will follow here.

The deep chain has a low magnet share yet almost never delivers the evidence turn, so magnet share must be read together with depth (W2-c). On W2-b, the raw fragment-pair count (title overlap) does discriminate on the public stores, 5 [2–8] at intake and 1 after curation, so we keep it; on the four shapes above it is 0 or 1 by construction. We add a content-level index: sibling roll-ups embedded, mean-centered per store and compared by cosine, with the threshold calibrated on split nodes (each node's roll-up cut in two gives a true same-topic pair). At the strict threshold, near-duplicate sibling pairs are 0 [0–3] at intake and 0 [0–2] after curation; at the lenient one, 11 [8–19] and 6 [2–7]. Where title discipline is off, one store counts 138 title pairs but 2 by content, so the revision reports both indices (Appendix D).

Q3, S2. On the 18 supersede probes, Gardener declined 1 of 18 in each of three runs, answered the other 17 correctly and cited no stale value. Flat declined 2, 2 and 0, answered the rest correctly and cited none. MemTree, Mem0 and A-MEM declined none, answered 5 to 9 of 18 correctly and cited a stale value in 8 to 11. MemTree has no update step, so its share of that gap is built in; Mem0 and A-MEM do update records and still cited 8 to 11. The declines in the submitted LongMemEval run were not abstentions but missing records: the writer's intent classifier, seeing its own earlier replies in the channel transcript, stopped recording, and every declined question had at most 5 records stored from about 47 sessions. The harness now delivers the haystack one user turn per message, without that transcript; MemTree, Mem0 and A-MEM ingest each session whole, their default usage. On LongMemEval knowledge-update questions, where the update is implicit, we grade the stale-citation rate among answered questions under one grader. It is Gardener 6.9% (4 of 58), MemTree 5.1% (3 of 59), A-MEM 13.3% (8 of 60), Mem0 32.6% (14 of 43) and the flat control 28.3% (15 of 53). Without a tree the same writer stored half as many records (median 86 versus 175 per question) and invalidated a quarter as many (22 versus 79 in total).

Q4 (Ethics Statement). Both sentences are true in different senses, and the first was written too broadly. "No real-user data" meant no data from anyone other than the authors; "derived from real usage" means that the workload is not synthetic. The facts are these. The 8 fixtures (51 questions, including the 18 supersede probes) are self-authored and synthetic and concern no real person. The 31-message workload is a work journal that one author kept through an AI-agent personal secretary during the study; no third-party users, participants or externally collected records are involved, so there was no human-subjects data collection. It contains no names, contact details or other identifiers of any third party in either language; organizations appear as placeholders and people only by role. The Ethics Statement now says exactly this, and the workload is not released. Since every current-backbone claim now rests on public corpora, this workload supports no claim beyond the historical observation on the retired backbone.

AuthorsOpenReview ↗

Corrections (15 Nov) to results in this thread, 1 of 2: pinned-reader F1 re-measured; no structure answers better than the flat store

Corrections (15 Nov). Four results in our replies to you change: the pinned-reader F1 withdrawn on 13 Nov, the LongMemEval counts and stale rates under W2 and Q3, the threshold count under Q1, and the replication under E1. This comment gives the first; our next comment in this thread gives the others. Terms: a pin injects one branch of the store into the reader's context for a question; injection recall is the share of questions whose evidence turn is inside the injected block. The same 586 records are stored in four shapes: a session tree (one node per session), a magnet node (every record in one node), a deep chain (each session nested under the previous one), and a flat store (no nodes).

  1. Pinned-reader token F1, re-measured. With the evidence branch pinned, no structure answers better than the flat store without pins. Median token F1 is 0.38 to 0.42 across the five pinned cells and 0.42 to 0.46 across the five cells without pins (flat store: 0.449). For the session tree and the deep chain, every pinned run lies below every run without pins. We withdraw the statement that pinned-reader F1 follows injection recall. The deterministic injection recall (67.9% under a session tree, 0.5% under a magnet node, 1.0% under a deep chain) is unchanged. The titles of our second replies to Reviewers DuiZ, rCfc, and 9kJE stated an advantage under injection, and we have edited them to remove it. On 13 Nov we edited four earlier comments in place to withdraw the F1 values first posted (0.361 for the session tree, 0.300 for the magnet node, and 0.265 for the deep chain, versus 0.312 for the flat store). In those runs, error replies returned while the model API was rate-limited had been graded as answers. From our second reply to you we also removed the fallback counts, the single-hop F1 values, and the retrieval calls per question, which came from the same runs. The values below replace them. We reran every cell with reader retries: three runs for each of the four shapes, and one run on each of the five stores that Gardener wrote and on their copies after curation.
cellrunstoken F1, median [min–max]questions with a fallback to global search, median [min–max]retrieval calls per questionunanswered questions, min–max
session tree, pinned30.397 [0.379–0.403]9 [9–22]0.956–8
magnet node, pinned30.407 [0.357–0.448]3 [3–27]3.078–12
deep chain, pinned30.383 [0.375–0.386]30 [25–32]1.2015–19
Gardener stores, pinned50.422 [0.367–0.448]3 [1–5]0.413–8
curated Gardener stores, pinned50.377 [0.354–0.438]5 [2–11]0.686–8
session tree, no pin30.456 [0.453–0.459]01.348–11
magnet node, no pin30.425 [0.399–0.455]01.779–12
deep chain, no pin30.458 [0.427–0.480]01.737–8
Gardener stores, no pin50.422 [0.370–0.462]00.773–8
flat store, no pin30.449 [0.408–0.494]01.214–8

A fallback to global search is a query that found nothing inside the pinned branch and was repeated over the whole store; the column counts the questions with at least one. Retrieval calls per question counts the reader's search calls over all attempts at a question (up to eight). Single-hop F1, which we had posted as higher for the pinned session tree, is 0.49 for both the pinned session tree and the flat store.

After retries, no remaining failure comes from a rate-limited call. A reply that stays empty after eight attempts, because the reader stopped at its step limit without writing an answer, is scored as wrong; this happens to 3 to 19 of 56 questions per run.

The revised PDF, which we will upload later today, reports these values in Section 6.6 and in Table 7 of Appendix H.

Reviewer_rCfcOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents Gardener, a hierarchical memory system for LLM agents that focuses on maintaining memory structure during continuous accumulation. It identifies three structural problems: magnet occupancy, fragmentation, and flatness. Gardener combines anti-preferential placement with online structural curation, including merging, splitting, and relocating existing records. The paper also introduces structural-health metrics to evaluate memory organization beyond conventional recall accuracy, and evaluates the system against existing memory approaches and across different LLM backbones.

Critical strengths
  1. The paper identifies concrete structural problems in hierarchical memory accumulation and proposes corresponding evaluation metrics. The problem formulation, proposed methods, and experimental analysis are well aligned, providing a useful perspective beyond conventional recall-based evaluation.
Critical weaknesses
  1. Limited distinction from existing memory management approaches. The paper does not clearly establish its novelty relative to existing memory systems. Prior work, such as xMemory and filesystem-based memory approaches, already supports structural reorganization through operations such as merging, splitting, and reassignment. The additional contributions of Gardener over these approaches remain unclear.

  2. Insufficient experimental evidence for the benefits of structural maintenance. The experimental evidence does not sufficiently support the paper's central claim that post-insertion structural maintenance provides meaningful benefits beyond existing placement strategies. In particular, Gardener achieves identical downstream results to Unstructured Flat, while the structural-health improvements are not clearly connected to downstream performance. The lack of controlled ablations isolating structural curation makes it difficult to establish the effectiveness of the main contribution.

Questions for authors

Q1. Could the authors clarify the advantages of Gardener over simpler unstructured memory approaches and existing systems supporting structural reorganization? Controlled comparisons on public benchmarks would help establish its contributions.

Q2. Could the authors provide ablations that isolate post-insertion curation from the placement policy, and examine how individual operations affect downstream performance?

Q3. What motivates the proposed structural-health metrics and their thresholds? Are they intended as diagnostics or optimization objectives? What benefits do LLM-driven curation operations provide over simpler rule-based or clustering-based alternatives?

7 comments
AuthorsOpenReview ↗

Reply to rCfc (1 of 2): novelty narrowed against xMemory; public-benchmark comparison, the tree store beats flat on LongMemEval; curation isolated before and after on five stores

Note: See our corrections of 15 Nov below.

We accept both weaknesses: the submission does not isolate curation from placement, and its only curation evidence is one run on a private corpus. As Q1 asks, we set the private corpus aside and rebuilt the evidence on public benchmarks during the discussion period. On LoCoMo conversation 2 (586 turns) we ran seven intake cells at five runs each with an exact permutation test, replicated the decisive cell on two further conversations, curated five stores over three passes to isolate curation, and measured branch-level injection on four store shapes. On the LongMemEval knowledge-update subset (72 questions) we ran Gardener, a flat control and a single-call reader over the Gardener store alongside MemTree, Mem0 and A-MEM, all under one grader, with stale citations post-graded among answered questions. Every number below comes from these runs. Terms: magnet share is the fraction of records in the largest node; unclassified records are those left at the root; fragment pairs are sibling nodes with near-duplicate titles; an ask-back is a placement question put to the user.

W1, Q1 (novelty). The distinction from xMemory is narrower than the introduction implies:

xMemoryGardener
Reorganization scopeperiodic repartition of the top theme layermerge, split, relocate at any depth
Source recordsrewrittenimmutable; an update marks the old record invalid
Who writesthe LLMthe LLM proposes, deterministic code validates and applies
Uncertaintynoneask the user or hold
Accumulation dynamicsnot measuredmeasured over the stream

The file-system memory study is cited as motivation for validated writes, not as a competing implementation. The fifteen surveyed systems are listed in a new appendix.

Q1 (public benchmarks). Current-backbone claims now rest on LoCoMo and LongMemEval only; the private corpus remains as the historical observation on the retired backbone. On LongMemEval knowledge-update questions, MemTree (our reimplementation), Mem0, A-MEM, flat and Gardener answer the same 72 questions, and answered replies are graded for stale citations by one grader: Gardener 6.9%, MemTree 5.1%, A-MEM 13.3%, Mem0 32.6% and the flat control 28.3%. Against the flat store with the same writer and reader, the tree answers 63 of 72 knowledge-update questions versus 41; without a tree the same writer stored half as many records and invalidated a quarter as many. MemTree's structure on the same 586-turn stream falls outside Gardener's range in every setting we tried. With its paper values (theta 0.4, lambda 0.5), three runs give a median magnet share of 0.915 on 3 to 4 nodes. With theta 0.3 it gives 0.017 on 165 nodes, 162 of them siblings at depth 2; with lambda 0.1 it gives 0.007 on 390 nodes in chains 10 deep. Gardener's baseline is 0.040 on 83 nodes at depth 3.

(Continued in part 2.)

AuthorsOpenReview ↗

Reply to rCfc (2 of 2): under branch-level injection, structure sets what is injected; thresholds are diagnostics, not controls

Note: See our corrections of 15 Nov below.

(Part 2 of 2; terms and conditions are in part 1.)

W2, Q2 (isolating curation). Curation is now isolated from placement by comparing each store before and after curation: the same five baseline stores, no re-ingestion. The submitted run reassigned 7 records and never triggered a merge or a split; on the public stores both trigger. Pass one reassigns a median of 10 records [4–16], merges 15 nodes [12–23], and splits 7 nodes [6–8]; fragment pairs fall from 5 to 1, unclassified share from 2.1% to 0.2%, and node count from 83 to 69. Magnet share rises from 0.040 to 0.059 because merges enlarge the largest node; it stays under the 0.30 gate. Splitting is capped at two nodes per pass, so three passes do not converge (magnet share 0.088 after pass three). The drop in unclassified share comes mostly from resolved ask-backs (turning ask-backs off alone gives 0.2%, p = 0.048), so the curation claim rests on the fragment-pair reduction, not on unclassified or magnet share. Per-operation downstream effects remain to be measured and are left for future work; the injection result below isolates structure as a whole.

W2 (flat parity). The harness reader searches all records and injects nothing by branch, so to it a tree and a flat store are the same store. Under branch-level injection, with the branch holding the evidence pinned for each question, the evidence turn reaches the model for 67.9% of 193 questions with a session tree, 0.5% with a magnet node and 1.0% with a deep chain (deterministic). Details are in our reply to Reviewer DuiZ. Correction (13 Nov): the pinned-reader F1 values first posted here (0.361, 0.300, 0.265 versus 0.312) are withdrawn; in those runs, 24 to 33 of the 56 reader replies per run were error replies returned while the model API was rate-limited, and our harness graded them as answers. The deterministic rates above are unaffected, and corrected F1 values will follow after re-running with retries.

Q3. The thresholds are diagnostics, not controls: they were fixed in July 2026 before the evaluation, no code reads them, and the stores built from 10 of 10 LoCoMo dialogues and 100 of 102 LongMemEval questions meet the same thresholds. We do not compare against a rule-only or clustering-only curator, so we do not claim that one would do worse. The LLM's role in curation is to propose merges and splits by meaning; the rules validate and apply them. The deterministic part of the write path is rule-based, and MemTree, which places by cosine similarity and never revisits a placement, serves as the clustering-based structural control.

AuthorsOpenReview ↗

Corrections (15 Nov) to results in this thread, 1 of 2: pinned-reader F1 re-measured; no structure answers better than the flat store

Corrections (15 Nov). Three results in our replies to you change: the pinned-reader F1 withdrawn on 13 Nov, the LongMemEval counts and stale rates under Q1, and the threshold count under Q3. This comment gives the first; our next comment in this thread gives the others. Terms: a pin injects one branch of the store into the reader's context for a question; injection recall is the share of questions whose evidence turn is inside the injected block. The same 586 records are stored in four shapes: a session tree (one node per session), a magnet node (every record in one node), a deep chain (each session nested under the previous one), and a flat store (no nodes).

  1. Pinned-reader token F1, re-measured. With the evidence branch pinned, no structure answers better than the flat store without pins. Median token F1 is 0.38 to 0.42 across the five pinned cells and 0.42 to 0.46 across the five cells without pins (flat store: 0.449). For the session tree and the deep chain, every pinned run lies below every run without pins. We withdraw the statement that pinned-reader F1 follows injection recall. The deterministic injection recall (67.9% under a session tree, 0.5% under a magnet node, 1.0% under a deep chain) is unchanged. The titles of our second replies to Reviewers DuiZ, rCfc, and 9kJE stated an advantage under injection, and we have edited them to remove it. On 13 Nov we edited four earlier comments in place to withdraw the F1 values first posted (0.361 for the session tree, 0.300 for the magnet node, and 0.265 for the deep chain, versus 0.312 for the flat store). In those runs, error replies returned while the model API was rate-limited had been graded as answers. From our reply to Reviewer DuiZ we also removed the fallback counts, the single-hop F1 values, and the retrieval calls per question, which came from the same runs. The values below replace them. We reran every cell with reader retries: three runs for each of the four shapes, and one run on each of the five stores that Gardener wrote and on their copies after curation.
cellrunstoken F1, median [min–max]questions with a fallback to global search, median [min–max]retrieval calls per questionunanswered questions, min–max
session tree, pinned30.397 [0.379–0.403]9 [9–22]0.956–8
magnet node, pinned30.407 [0.357–0.448]3 [3–27]3.078–12
deep chain, pinned30.383 [0.375–0.386]30 [25–32]1.2015–19
Gardener stores, pinned50.422 [0.367–0.448]3 [1–5]0.413–8
curated Gardener stores, pinned50.377 [0.354–0.438]5 [2–11]0.686–8
session tree, no pin30.456 [0.453–0.459]01.348–11
magnet node, no pin30.425 [0.399–0.455]01.779–12
deep chain, no pin30.458 [0.427–0.480]01.737–8
Gardener stores, no pin50.422 [0.370–0.462]00.773–8
flat store, no pin30.449 [0.408–0.494]01.214–8

A fallback to global search is a query that found nothing inside the pinned branch and was repeated over the whole store; the column counts the questions with at least one. Retrieval calls per question counts the reader's search calls over all attempts at a question (up to eight). Single-hop F1, which we had posted as higher for the pinned session tree, is 0.49 for both the pinned session tree and the flat store.

After retries, no remaining failure comes from a rate-limited call. A reply that stays empty after eight attempts, because the reader stopped at its step limit without writing an answer, is scored as wrong; this happens to 3 to 19 of 56 questions per run.

The revised PDF, which we will upload later today, reports these values in Section 6.6 and in Table 7 of Appendix H.

Reviewer_9kJEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies long-term memory for large language model agents, or LLM agents, and argues that existing online systems only decide where to put a new record and never revisit what was placed before. On a 31-message multi-topic workload, the authors report that undisciplined insertion creates a magnet node holding 64% of all records, near-duplicate sibling nodes, and a flat tree. They propose Gardener, a tree memory with three parts: anti-preferential placement, where deterministic code scores candidate nodes with per-keyword voting and square-root size normalization and the LLM only chooses among them; an online curation loop that merges, splits, relocates and renames nodes under deterministic guards; and a three-way triage that places a record, asks the user, or holds it. Value updates are non-destructive: the old record is marked invalid instead of deleted. They also propose four structural-health metrics, magnet occupancy, fragmentation, depth distribution and unclassified rate, plus a writer attribution that compares an oracle placement cell with the end-to-end pipeline. On a 51-question synthetic suite with gemini-flash-latest, Gardener reaches 84.3% recall, the same as a flat control and close to MemTree, Mem0 and A-MEM, and 0% stale-value citations against about 50% for the baselines. On the structural axis, Gardener has 31% magnet occupancy at intake and 22% after one curation pass, against 42% for a MemTree reimplementation. Ablations on two backbones show that removing all guards collapses the structure to 64% on the original backbone but only to 33% on the current one.

Critical strengths

The question is real and under-studied. The related work survey is careful, and the point that recall benchmarks cannot see structural degradation is well made. The non-destructive supersede mechanism with deterministic write guards gives a clean and reproducible result: 0% stale-value citations in every run, and original-value temporal queries still answerable, which destructive updates cannot do. I also appreciate that the authors report inconvenient facts openly, including the backbone being retired mid-study, the flat control matching Gardener on recall and freshness, the negative writer gap on the original backbone, and the 15.7% response avoidance counted as failure. Ablations are run on two backbones and the harness and the MemTree reimplementation are released.

Critical weaknesses

The primary experiment does not support the central claim. The thesis is that a placement discipline plus a maintenance loop are needed to keep the structure healthy. On the primary English evaluation with the current backbone, Figure 2 shows that removing all six guards changes the magnet metric from 31% to 33%, and removing the anti-magnet scoring alone leaves it at 31%. The only evidence that anti-preferential placement matters is the 64% collapse on the retired Gemini 2.5 Flash backbone on a Korean corpus that is withheld. The paper calls this a difference in backbone temperament, but to me it means the headline mechanism has no measurable effect in the headline setting. The curation result is one pass on 31 messages that reassigns 7 records, moving magnet from 31% to 22% and unfiled from 33% to 4%. The 33% unfiled rate is produced by the authors’ own hold and ask-back policy, so the curation loop is partly repairing a problem that the triage itself creates. This is the main reason for my rating.

The structural-health metrics are never connected to any downstream cost. The paper states that a magnet erodes branch-level context injection, summary coherence and navigability, but none of these is measured anywhere. The one downstream quantity that is measured, recall, shows parity: Gardener end-to-end is 84.3%, exactly the flat control with no tree, and the 0% stale citations also come from the flat control. So on the primary suite the tree contributes nothing measurable, and the paper itself says the freshness result comes from the supersede guard, not from the structure. The gate thresholds, 0.30, 0 and 0.10, are described as regression cut lines derived from the same corpus and the same system that are then evaluated against them, which is circular. The fragmentation index is 0 for every system in every table, so it never discriminates between anything.

The scale and the comparison conditions are too thin for the numbers to carry the conclusions. The structural workload is 31 messages from one user in one domain; a magnet difference of 31% against 42% is about three records. Ablation cells on the current backbone are single runs. The fixture suite is 51 questions authored by the authors with LLM assistance, and the 18 forbidden questions target exactly the supersede behavior that Gardener implements, so the 50% stale rate for baselines is partly by construction; on the original LongMemEval knowledge-update subset, where updates are implicit, Gardener scores 38 of 72. MemTree is a reimplementation without official code. The external baselines are queried by single-call injection while Gardener uses a tool-using reader with a multi-step active recall loop, so recall differences cannot be attributed to the memory rather than the reader. The internal prompts and all stored summaries are Korean while the primary workload is English, which is an odd asymmetry for a primary evaluation.

On presentation, the paper is hard to follow. Many numbers appear without the context needed to read them, terms such as QNA corpus, e8, oracle versus end-to-end and the pending queue are used before they are explained, Appendices C and D are empty headings, and the fifteen surveyed systems are never listed. A good amount of space goes to narration about the thesis rather than to evidence for it.

Questions for authors

On the current backbone, is there any setting in which the anti-preferential scoring changes the outcome? Figure 2 shows 31% with and without it. If the guards only matter on a retired backbone, what is the claim for a reader who uses a current model?

What downstream task does magnet occupancy actually hurt? Can you show one measured quantity, for example branch-level context injection or summary coherence, where a 64% magnet and a 20% magnet give different accuracy?

How were the thresholds 0.30, 0 and 0.10 chosen before the evaluation, and do they transfer to any corpus other than the one they were fit on?

Can Mem0, A-MEM and MemTree be run with the same tool-using reader and active recall loop as Gardener, so that recall and stale-citation differences are attributed to the memory and not to the reader?

The ask-back rate on the current backbone is 9 of 31 records, so the user is asked where to file about 29% of what they write. How should a reader interpret this cost, and what happens to magnet and unclassified rate if ask-back is disabled and those records are placed by the best candidate?

5 comments
AuthorsOpenReview ↗

Reply to 9kJE (1 of 2): presentation fixed; on public data one guard off collapses the tree in five of five runs; the 64% collapse confined to the retired backbone

Note: See our corrections of 15 Nov below.

We agree with the presentation points in full and with the scale point: n = 1 ablations on 31 records support no conclusion. On public data the central mechanism does have a measurable effect: with the hierarchy nudge off, a 586-turn conversation collapses to 7 nodes in five of five runs. To meet the scale point, we set the private corpus aside and rebuilt the evidence on public benchmarks during the discussion period. On LoCoMo conversation 2 (586 turns) we ran seven intake cells at five runs each with an exact permutation test, replicated the decisive cell on two further conversations, curated five stores over three passes, and measured branch-level injection on four store shapes. On the LongMemEval knowledge-update subset (72 questions) we ran Gardener, a flat control and a single-call reader over the Gardener store alongside MemTree, Mem0 and A-MEM, all under one grader (which also answers W3-d), with stale citations post-graded among answered questions. Every number below comes from these runs; the one private-corpus comparison (W1-b) is marked as such. Terms: magnet share is the fraction of records in the largest node; unclassified records are those left at the root; fragment pairs are sibling nodes with near-duplicate titles; an ask-back is a placement question put to the user. The hierarchy nudge asks the writer to open a sub-branch once a branch holds more than 8 direct records, and title discipline forbids sentence-like branch titles.

Presentation. Appendices C and D each contained a single figure that LaTeX placed on the following page, leaving the heading alone; both are now sections with text. The QNA corpus, e8, oracle versus end-to-end, and the pending queue are defined at first use. The fifteen surveyed systems are listed in a new appendix, the thesis sentence that Sections 1 and 5 both carried now appears in Section 1 only, apart from the conclusion, and single-run cells are marked as such (Table 2, Section 6.4). Every table caption now states the corpus and the number of runs behind its numbers.

W1-a, Q1. Two things made the submitted ablation uninformative. The three baseline runs alone spanned 27 to 55% magnet share, which covers every single-run ablation value. And sibling merge and bloat split (two of the six guards) run only during curation, so ablating them during ingestion could not show anything. On public data, two cells collapse the tree and the other two match the baseline:

cell (n = 5)magnet median [range]p vs baselinenodes
baseline0.040 [0.032–0.050]83
all four intake guards off1.00 [0.73–1.00]0.0486
hierarchy nudge off0.53 [0.47–0.54]0.0487
scoring term (anti-preferential) off0.042 [0.034–0.059]1.086
ask-back off0.040 [0.030–0.049]1.085

p = 0.048 is the smallest value an exact permutation test can give with five runs per cell. Turning off the hierarchy nudge alone reproduces the collapse, and three runs each on conversations 3 and 6 replicate it (magnet share 0.48 versus 0.044 and 0.96 versus 0.034, every nudge-off run above every baseline run). We narrow contribution (1): the paper now claims volume-blind, size-penalized scoring with deterministic validation of every write, and attributes the prevention of collapse to one size-triggered guard (code flags a branch with more than 8 direct records); contributions (2) to (4) stand. If that guard is lost, the failure takes one known form, and the magnet diagnostic detects it.

W1-b, W3-e. The 64% collapse exists only on the retired backbone and is removed from every current-backbone claim. On the private workload with the same backbone, the Korean and English versions behave alike (oracle recall 98.0% versus 96.1%, end-to-end recall 90.2% versus 84.3%, magnet share 35% versus 31%). The guard and scoring code does not branch on language. The system prompts stay in Korean even for English input; this comes from the deployment and is disclosed in Section 7.

(Continued in part 2.)

AuthorsOpenReview ↗

Reply to 9kJE (2 of 2): pending ask-backs explain the unclassified share; the tree beats flat on LongMemEval, and structure sets injection recall; stale rates for six systems; the equal-reader control

Note: See our corrections of 15 Nov below.

(Part 2 of 2; terms and conditions are in part 1.)

W1-c, Q5. The reviewer is right: the harness never answers a placement question, so every record that triggers one stays unclassified. With a simulated user who accepts the first candidate, magnet share is 0.043, unclassified 0.2% and nodes 80; with ask-backs off, magnet share 0.040, unclassified 0.2% and nodes 85; with questions left pending (the baseline), magnet share 0.040, unclassified 2.1% and nodes 83. Pending ask-backs therefore account for the unclassified share, and answering them does not change the structure. On the public stream the ask-back rate is 2%: a median of 12 questions over 586 turns. On five public stores, pass one of curation merges a median of 15 nodes [12–23] and splits 7 [6–8], operations the submitted single pass never triggered. The curation claim rests on the fragment-pair reduction (5 to 1 over those stores), not on magnet share, which rises after merges (0.040 to 0.059).

W2-a, Q2. The harness reader searches all records and injects nothing by branch, so to it a tree and a flat store are the same store. We therefore measured branch-level injection, where each node contributes its roll-up, a digest capped at 20 lines. The same 586 records were stored as a session tree (one node per session), one magnet node (every record in one node) and a deep chain (each session nested under the previous one), and for each question the branch holding the evidence was pinned (injected). The evidence turn reaches the model for 67.9%, 0.5% and 1.0% of 193 questions (deterministic). Correction (13 Nov): the pinned-reader F1 values (0.361, 0.300, 0.265 versus 0.312) and the fallback observation first posted here are withdrawn, because rate-limit error replies were graded as answers; corrected values will follow. On LongMemEval, with the same writer and reader, the tree store answers 63 of 72 knowledge-update questions and the flat store 41 (W3 below). Structure pays off where context is injected and where the writer places into it.

W2-b, Q3. The thresholds are diagnostics, not controls: they were fixed in July 2026 before the evaluation, no code reads them, and the stores built from 10 of 10 LoCoMo dialogues and 100 of 102 LongMemEval questions meet the same thresholds. We remove the phrase "regression cut lines". W2-c. We keep the raw fragment-pair count and add a content-level index: sibling roll-ups embedded, mean-centered per store and compared by cosine, with the threshold calibrated on split nodes (each node's roll-up cut in two gives a true same-topic pair). On the public stores it gives 11 [8–19] near-duplicate sibling pairs at intake and 6 [2–7] after curation at the lenient threshold, 0 [0–3] and 0 [0–2] at the strict one. It corrects the raw count where titles are long (138 versus 2 on one store with title discipline off). Both appear in Appendix D.

W3-a/b/c. Agreed. Because the 18 probes target supersede by construction, we also grade LongMemEval knowledge-update replies, where the update is implicit, with one grader and among answered questions only. The rates are Gardener 6.9% stale (4 of 58), MemTree 5.1% (3 of 59), A-MEM 13.3% (8 of 60), Mem0 32.6% (14 of 43) and the flat control 28.3% (15 of 53). Without a tree the same writer stored half as many records (median 86 versus 175 per question) and invalidated a quarter as many (22 versus 79 in total). The declines in the submitted run were not abstentions but missing records: the writer's intent classifier, seeing the secretary's own earlier replies in the transcript, stopped recording, and every declined question had at most 5 records stored from about 47 sessions. The harness now delivers the haystack one user turn per message, without that transcript; MemTree, Mem0 and A-MEM ingest each session whole, their default usage. On the 18 probes, Gardener declined 1 of 18 in each of three runs, answered the other 17 correctly and cited no stale value. MemTree, Mem0 and A-MEM declined none, answered 5 to 9 of 18 correctly and cited a stale value in 8 to 11. MemTree remains a reimplementation, with code released.

W3-d, Q4. Giving the external systems our tool-using reader is not feasible in this window, so we do not attribute recall differences to the memory design. The opposite control reads Gardener's store with the single-call responder the baselines use (hybrid top-10 retrieval, one call). By the same judge it answers 49 of 72 knowledge-update questions, versus 63 with the tool-using reader, 65 for MemTree, 59 for A-MEM and 37 for Mem0; its stale rate among answered questions is 10.4% (5 of 48). Under an equal reader the Gardener store therefore trails MemTree and A-MEM on recall, and the tool-using reader accounts for 14 questions. Invalidated records are filtered out in the store query, so the remaining stale citations come from updates the writer did not mark invalid, not from the reader.

AuthorsOpenReview ↗

Corrections (15 Nov) to results in this thread, 1 of 2: pinned-reader F1 re-measured; no structure answers better than the flat store

Corrections (15 Nov). Four results in our replies to you change: the pinned-reader F1 withdrawn on 13 Nov, the LongMemEval counts and stale rates under W2 and W3, the threshold count under W2-b, and the replication under W1-a. This comment gives the first; our next comment in this thread gives the others. Terms: a pin injects one branch of the store into the reader's context for a question; injection recall is the share of questions whose evidence turn is inside the injected block. The same 586 records are stored in four shapes: a session tree (one node per session), a magnet node (every record in one node), a deep chain (each session nested under the previous one), and a flat store (no nodes).

  1. Pinned-reader token F1, re-measured. With the evidence branch pinned, no structure answers better than the flat store without pins. Median token F1 is 0.38 to 0.42 across the five pinned cells and 0.42 to 0.46 across the five cells without pins (flat store: 0.449). For the session tree and the deep chain, every pinned run lies below every run without pins. We withdraw the statement that pinned-reader F1 follows injection recall. The deterministic injection recall (67.9% under a session tree, 0.5% under a magnet node, 1.0% under a deep chain) is unchanged. The titles of our second replies to Reviewers DuiZ, rCfc, and 9kJE stated an advantage under injection, and we have edited them to remove it. On 13 Nov we edited four earlier comments in place to withdraw the F1 values first posted (0.361 for the session tree, 0.300 for the magnet node, and 0.265 for the deep chain, versus 0.312 for the flat store). In those runs, error replies returned while the model API was rate-limited had been graded as answers. From our reply to Reviewer DuiZ we also removed the fallback counts, the single-hop F1 values, and the retrieval calls per question, which came from the same runs. The values below replace them. We reran every cell with reader retries: three runs for each of the four shapes, and one run on each of the five stores that Gardener wrote and on their copies after curation.
cellrunstoken F1, median [min–max]questions with a fallback to global search, median [min–max]retrieval calls per questionunanswered questions, min–max
session tree, pinned30.397 [0.379–0.403]9 [9–22]0.956–8
magnet node, pinned30.407 [0.357–0.448]3 [3–27]3.078–12
deep chain, pinned30.383 [0.375–0.386]30 [25–32]1.2015–19
Gardener stores, pinned50.422 [0.367–0.448]3 [1–5]0.413–8
curated Gardener stores, pinned50.377 [0.354–0.438]5 [2–11]0.686–8
session tree, no pin30.456 [0.453–0.459]01.348–11
magnet node, no pin30.425 [0.399–0.455]01.779–12
deep chain, no pin30.458 [0.427–0.480]01.737–8
Gardener stores, no pin50.422 [0.370–0.462]00.773–8
flat store, no pin30.449 [0.408–0.494]01.214–8

A fallback to global search is a query that found nothing inside the pinned branch and was repeated over the whole store; the column counts the questions with at least one. Retrieval calls per question counts the reader's search calls over all attempts at a question (up to eight). Single-hop F1, which we had posted as higher for the pinned session tree, is 0.49 for both the pinned session tree and the flat store.

After retries, no remaining failure comes from a rate-limited call. A reply that stays empty after eight attempts, because the reader stopped at its step limit without writing an answer, is scored as wrong; this happens to 3 to 19 of 56 questions per run.

The revised PDF, which we will upload later today, reports these values in Section 6.6 and in Table 7 of Appendix H.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Corrections, public-corpus re-measurement at n = 5, and a revised PDF by Nov 15: one guard off collapses the tree; the tree store beats flat on LongMemEval

Note: See our corrections of 15 Nov below.

Response to all reviewers

We thank the reviewers. All three accept the framing, that recall benchmarks miss structural degradation, and two accept the supersede result. All three find the evidence for the intake guards and for the value of the tree insufficient, and all three note that the curation evidence is a single pass on a private corpus. We agree with these findings. Our response has four parts: corrections, public-corpus measurements, the resulting changes to our claims, and the revision.

1. Corrections. The manuscript says in five places (abstract, Sections 6.2 and 6.5, conclusion, Appendix E) that guarded intake stays within the health gate. It does not: at intake, magnet share is 31% and unclassified share 33%, and only the post-curation state (22% and 4%) meets the gate. The code imposes no cap on a node's share of records; Equation 1 bounds a candidate's score. The curated column of Table 2 is one manual pass, not a median. We withdraw the claim that the guards are model-independent. The Ethics Statement is rewritten: the fixtures are synthetic, and the 31-message workload derives from the authors' own use, contains no third-party identifiers, and is not released.

2. Public-corpus measurements. We set the private corpus aside. Every current-backbone claim about structure and freshness is re-measured from scratch on public benchmarks: LoCoMo (conversation 2, 586 turns, five runs per cell, exact permutation test) and on the LongMemEval knowledge-update subset (72 questions), with flat, MemTree, Mem0 and A-MEM under one grader. Terms: magnet share is the fraction of records in the largest node; fragment pairs are sibling nodes with near-duplicate titles; the hierarchy nudge asks the writer to open a sub-branch once a branch holds more than 8 direct records. In the table, p = 0.048 is the smallest value an exact permutation test can give with five runs per cell.

MeasurementResult
Intake ablation (LoCoMo, n = 5)all four guards off: magnet share 1.00, 6 nodes (p = 0.048); hierarchy nudge off alone: 0.53, 7 nodes (p = 0.048); scoring term off: 0.042, 86 nodes (p = 1.0); baseline 0.040, 83 nodes
Replication (conversations 3 and 6, n = 3)nudge off 0.48 versus 0.044 and 0.96 versus 0.034, every nudge-off run above every baseline run
Curation (five stores, three passes)fragment pairs 5 to 1, unclassified 2.1% to 0.2%, magnet share 0.040 to 0.059 (merges enlarge the largest node)
Branch-level injection (193 questions)evidence turn inside the injected block: session tree 67.9%, magnet node 0.5%, deep chain 1.0%; pinned-reader F1 withdrawn and being re-measured (see correction below)
LongMemEval stale citations among answeredGardener 6.9%, MemTree 5.1%, A-MEM 13.3%, Mem0 32.6%, flat control 28.3%
LongMemEval correct of 72 (one judge)Gardener 63, MemTree 65, A-MEM 59, flat 41, Mem0 37; Gardener store read by the baselines' single-call reader 49

3. What changes in our claims. Contribution (1) narrows: the paper now claims volume-blind, size-penalized scoring with deterministic validation of every write, and attributes the prevention of collapse to one size-triggered guard, the hierarchy nudge; the scoring term alone has no measurable effect. The curation claim rests on the fragment-pair reduction, not on magnet share. The unclassified share is explained by pending ask-backs: turning them off gives 0.2%. The tree's value appears where context is injected by branch and where the writer places into it. With the same writer and reader, the tree store answers 63 of 72 LongMemEval knowledge updates and the flat store 41, because without a tree the writer stored half as many records and invalidated a quarter as many. Under an equal single-call reader the Gardener store trails MemTree and A-MEM on recall; we report this as is and do not attribute recall differences to the memory design. The private 31-record corpus remains only as the historical observation on the retired backbone.

4. Revision. A revised PDF (body 10 pages) with these corrections, a new Section 6.6, and new appendices (public-corpus protocol and tables, the fifteen surveyed systems) will be uploaded with a summary of changes by November 15. Reviewer-specific details, including the supersede probes, the xMemory comparison, and the equal-reader control, are in the three threads below.

Correction (13 Nov). The pinned-reader F1 values first posted in this thread (0.361, 0.300, 0.265 versus 0.312) are withdrawn. In those runs, between 24 and 33 of the 56 reader replies per run were error replies ("failed to generate a response") returned while the model API was rate-limited, and our harness graded them as answers. The deterministic injection rates do not involve the reader and are unaffected. These runs are being repeated with retries; corrected values will follow in the reviewer threads and the revision.

AuthorsOpenReview ↗

Corrections (15 Nov), 1 of 2: pinned-reader F1 re-measured; no structure answers better than the flat store

Corrections (15 Nov). Four results posted on 13 Nov change: two were wrong, one was withdrawn on 13 Nov and is now re-measured, and one now has five runs per cell. Each reviewer thread carries the same note. This comment gives the first; our next comment in this thread gives the others. Terms: a pin injects one branch of the store into the reader's context for a question; injection recall is the share of questions whose evidence turn is inside the injected block. The same 586 records are stored in four shapes: a session tree (one node per session), a magnet node (every record in one node), a deep chain (each session nested under the previous one), and a flat store (no nodes).

  1. Pinned-reader token F1, re-measured. With the evidence branch pinned, no structure answers better than the flat store without pins. Median token F1 is 0.38 to 0.42 across the five pinned cells and 0.42 to 0.46 across the five cells without pins (flat store: 0.449). For the session tree and the deep chain, every pinned run lies below every run without pins. We withdraw the statement that pinned-reader F1 follows injection recall. The deterministic injection recall (67.9% under a session tree, 0.5% under a magnet node, 1.0% under a deep chain) is unchanged. The titles of our second replies to Reviewers DuiZ, rCfc, and 9kJE stated an advantage under injection, and we have edited them to remove it. On 13 Nov we edited four earlier comments in place to withdraw the F1 values first posted (0.361 for the session tree, 0.300 for the magnet node, and 0.265 for the deep chain, versus 0.312 for the flat store). In those runs, error replies returned while the model API was rate-limited had been graded as answers. From our reply to Reviewer DuiZ we also removed the fallback counts, the single-hop F1 values, and the retrieval calls per question, which came from the same runs. The values below replace them. We reran every cell with reader retries: three runs for each of the four shapes, and one run on each of the five stores that Gardener wrote and on their copies after curation.
cellrunstoken F1, median [min–max]questions with a fallback to global search, median [min–max]retrieval calls per questionunanswered questions, min–max
session tree, pinned30.397 [0.379–0.403]9 [9–22]0.956–8
magnet node, pinned30.407 [0.357–0.448]3 [3–27]3.078–12
deep chain, pinned30.383 [0.375–0.386]30 [25–32]1.2015–19
Gardener stores, pinned50.422 [0.367–0.448]3 [1–5]0.413–8
curated Gardener stores, pinned50.377 [0.354–0.438]5 [2–11]0.686–8
session tree, no pin30.456 [0.453–0.459]01.348–11
magnet node, no pin30.425 [0.399–0.455]01.779–12
deep chain, no pin30.458 [0.427–0.480]01.737–8
Gardener stores, no pin50.422 [0.370–0.462]00.773–8
flat store, no pin30.449 [0.408–0.494]01.214–8

A fallback to global search is a query that found nothing inside the pinned branch and was repeated over the whole store; the column counts the questions with at least one. Retrieval calls per question counts the reader's search calls over all attempts at a question (up to eight). Single-hop F1, which we had posted as higher for the pinned session tree, is 0.49 for both the pinned session tree and the flat store.

After retries, no remaining failure comes from a rate-limited call. A reply that stays empty after eight attempts, because the reader stopped at its step limit without writing an answer, is scored as wrong; this happens to 3 to 19 of 56 questions per run.

The revised PDF, which we will upload later today, reports these values in Section 6.6 and in Table 7 of Appendix H.

AuthorsOpenReview ↗

Corrections (15 Nov), 2 of 2: LongMemEval accuracy 55 of 72, not 63; the zero-fragment-pair threshold holds in 2 of 36 stores before curation

Corrections (15 Nov), continued. This comment gives the remaining corrections. Terms: magnet share is the fraction of records in the largest node; unclassified records are those left at the root; fragment pairs are sibling nodes with near-duplicate titles; a stale citation is an answer that states a value the conversation later replaced. The hierarchy nudge asks the writer to open a sub-branch once a branch holds more than 8 direct records.

  1. LongMemEval knowledge updates, reader rerun (correction). The reader of the Gardener and flat runs had also run while the model API was rate-limited, so we reran it on the same 156 stores. Gardener now answers 55 of 72 questions correctly (posted: 63), and the flat store answers 45 (posted: 41). Among answered questions, Gardener cites a stale value in 8.8% of replies (5 of 57; posted: 6.9%), and the flat store in 24.6% (14 of 57; posted: 28.3%). Between the two reader runs, 17% to 19% of the questions change their grade for each of the two systems, so single runs that differ by a few questions are within reader variance. Gardener still cites stale values about a third as often as the flat store. With its own reader, Gardener now also answers fewer questions correctly than MemTree (65 of 72) and A-MEM (59 of 72), as it already did under the equal single-call reader.

  2. Threshold count (correction). We wrote that the stores built from 10 of 10 LoCoMo dialogues meet the three thresholds (magnet share at most 0.30, unclassified share at most 0.10, zero fragment pairs). Re-measured under the current protocol before any curation, on three to five stores per dialogue (36 stores), the magnet and unclassified thresholds hold in every store: magnet share is at most 0.16 and the unclassified share at most 0.05. The zero-fragment-pair threshold holds in 2 of the 36 stores and at the median of no dialogue; the per-dialogue medians are 2 to 10 fragment pairs. Three curation passes bring conversation 2 to 0 or 1 pairs in each of its five stores. The LongMemEval count, 100 of 102 stores (one per question), refers to the magnet threshold only.

  3. Replication, now with five runs per cell. With the hierarchy nudge off, median magnet share is 0.59 versus 0.054 for the baseline on conversation 3, and 0.53 versus 0.036 on conversation 6 (posted with three runs per cell: 0.48 versus 0.044, and 0.96 versus 0.034). Every nudge-off run lies above every baseline run, and the exact permutation p is 0.048 on each conversation, the smallest value five runs per cell allow.

The revised PDF, which we will upload later today, reports these values in Section 3.3, Section 6.6, and Appendix H (Tables 5 and 8).

AuthorsOpenReview ↗

Summary of changes (revision uploaded 15 Nov)

Summary of changes (revision uploaded 15 Nov)

The revision corrects the claims listed below. It adds one subsection (6.6) and one appendix (H) with the public-corpus measurements described in our replies and in our corrections of 15 Nov, and an appendix (I) that lists the surveyed systems. Body length is 10 pages; no claim was broadened.

WhereChange
AbstractWe remove "guarded accumulation stays within the health gate" and state the public-corpus result: with the guards off, a 586-turn LoCoMo conversation collapses to one node, and the hierarchy nudge accounts for the collapse. We label the 64% collapse as an original-backbone result. We replace "eliminate" with the measured rates: 0% on the probes and 8.8% of answered LongMemEval knowledge updates, versus 24.6% for the flat control and 5.1% to 32.6% for the three baselines. We state that reader accuracy did not follow injection recall.
Section 1We reword contribution (1) to volume-blind, size-penalized scoring with deterministic validation of every write, and point to Section 6.6 for the one guard that prevents the collapse. Contribution (4) no longer says that the comparisons show a gap in existing systems.
Section 3.3We state that the thresholds were fixed in July 2026 before the evaluation and that no code reads them, remove "regression cut lines", and add the transfer evidence measured before curation: the magnet and unclassified thresholds hold in all 36 stores built from the 10 LoCoMo dialogues, the zero-fragment-pair threshold in 2 of the 36, and the magnet threshold in 100 of 102 LongMemEval stores.
Section 4We state that neither scoring device caps a node's share of records and describe the score as volume-blind and size-penalized. We define the terms of Equation 1 and write its title term per keyword, as in the code. We define roll-up, pending queue, and forbidden answer at first use, and state the nudge threshold as 8 direct records.
Section 5We restate the downstream cost of the pathologies as a hypothesis and test it in Section 6.6.
Section 6.1, Table 2We bound the statement that freshness comes from the write guard and not from the tree to the probes, and add the LongMemEval contrast (flat control 24.6% stale versus 8.8%). We mark the curated column of Table 2 as a single run.
Section 6.2We correct the claim: intake on the current backbone does not meet the gate (magnet share 31%, unclassified 33%), and the 64% collapse is attributed to the original backbone only.
Section 6.3, Appendix AWe narrow "all ten dialogues meet the gates" to the magnet and unclassified thresholds; the fragment-pair threshold does not transfer.
Section 6.4We retitle it "Ablations on the private workload" and state that the current-backbone n = 1 cells support no conclusion.
Section 6.5We withdraw the model-independence claim and state the effect per backbone.
Section 6.6 (new)We add the public-corpus re-measurement: the LoCoMo ablation at n = 5 with an exact permutation test (Table 3), replication on two conversations at n = 5, curation variance, injection recall under branch-level injection with the reader result (pinned-reader F1 does not follow injection recall), and LongMemEval stale rates.
Discussion, ConclusionWe align both with the above and add the limitation that no accuracy gain from structural health has been shown. We replace "the two languages behave equivalently" with the measured difference.
Reproducibility, EthicsThe fixtures are synthetic; the 31-message workload derives from the authors' own use, contains no third-party identifiers, and is not released. The supplementary package now includes the scripts and result files of the public-corpus measurements (file map in results/README.md). The supplementary figure script matches the revised figures.
Appendix C, DWe fold the two figure-only appendices into sections with text, add the public curation table (Table 4, five stores each for conversations 2, 3, and 6), and add an embedding-based sibling index beside the title-token count.
Appendix EWe correct the portability conclusion: two of three backbones meet the gate at intake.
Appendix H (new)We add the public-corpus protocol and tables: replication at n = 5 on conversations 3 and 6, with four further ablation cells at n = 3 (Table 5); MemTree on the same LoCoMo stream (Table 6); injection recall and the re-measured reader F1 for ten cells (Table 7); and LongMemEval stale rates for Gardener, a single-call reader over its store, the flat control, MemTree, A-MEM and Mem0 (Table 8), with the Gardener and flat rows from a second reader run.
Appendix I (new)We list the fifteen surveyed systems and whether each revisits placed records.
ThroughoutWe use one name per term, remove narrative sentences, and split long sentences.

Reviewer-specific details are in the three threads below.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.