acceptodds
Under review as a conference paper at ICLR 2027

Cognifold: Always-On Proactive Memory via Cognitive Folding

Abstract

Agents receive fragmented observations over time, creating a need to organize experience into reusable memory. Existing agent memory remains predominantly reactive and retrieval-based, lacking the capacity to autonomously organize events into persistent cognitive structure. Toward genuinely autonomous agents, we introduce , a proactive agent memory that continuously folds fragmented event streams into persistent structures, bootstrapping progressively higher-level cognition from incoming events and accumulated knowledge. organizes memory into source events, reusable concepts, and candidate intents. As the stream arrives, cognitive structures are proactively assembled, interpreted through accumulated structure, merged when semantically similar, and reweighted by recency, while candidate intents surface from converging concept and event evidence. We evaluate structure formation with CogEval-Bench, a benchmark whose ground truth is established by construction, and show that uniquely produces event-grounded concepts that match the reference structure together with supported candidate intents. also performs robustly on conventional long-term memory benchmarks (LoCoMo, LongMemEval).

Primary area:
foundation or frontier models, including LLMs
Keywords:
Proactive Agents, LLM Memory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1SWROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CogniFold, which organizes continuously arriving information into three layers: events, concepts, and candidate intents. When a new event arrives, the system retrieves existing concepts and uses an LLM to generate a structured update plan, prioritizing the reuse or revision of existing concepts while preserving provenance links.

Critical strengths

For continuously operating agents, how to develop reusable understanding as information arrives, rather than merely storing historical records, is a worthwhile research problem.

Critical weaknesses

1.In the introduction, the authors characterize existing memory architectures as having fixed topologies once formed. The related work section also describes A-Mem as structurally static. In fact, A-Mem explicitly includes dynamic link creation triggered by new memories and updates to existing memory representations. Using existing memories to interpret new information, dynamically modifying memory content and connections, reusing or merging concepts across events, and designing suitable memory structures and systems cannot, in themselves, be considered contributions unique to this paper. The focus of this work is its explicit event–concept–intent organization and a memory-writing process designed around concept reuse. The issue is that the paper has not sufficiently explained which mechanisms enable capabilities that existing methods would struggle to achieve. What advantages does the proposed method offer over related dynamic memory methods under the same experimental conditions?

2.Based on the evaluation procedure described in the main text, the workflow can be summarized as follows: a new event arrives, similar existing concepts are retrieved, an LLM decides whether to update an existing concept or create a new one, and the results are stored. My question is whether a conventional memory store that maintains concept text and supports retrieval and modification could do the same. In LoCoMo, the system retrieves the 15 most similar existing concepts and provides their IDs and descriptions to the model. It therefore remains unclear how much the connections, hierarchy, and other topological information among concepts actually contribute.

Questions for authors

Please see the specific concerns and questions raised in the Weaknesses section.

2 comments

Thank you for the clear and focused review. Both weaknesses point to important issues in how we described prior dynamic memory systems and what specific role the graph structure plays. We have revised Sections 1, 2, and 4.5 (marked in blue) and added controlled ablations addressing both points.

1. Positioning Relative to A-Mem, Proactive Organization, and Same-Model Comparisons (W1)

CogniFold goes beyond generic ingestion-time updates to perform proactive organization that continuously tracks user intents. A-Mem dynamically links notes and updates note representations on arrival, and Mem0, Zep, EverMemOS, and PlugMem likewise update memory during ingestion (credited in Sections 1 and 2). CogniFold differs in two ways: (i) instead of linking notes or indexing facts without user goals and passively awaiting queries, it performs proactive organization, folding recurring observations around user concerns to track grounded intents ahead of queries (85.4% valid and 43.1% realized 3 to 4.5 sessions later across 176 LoCoMo intents; Table 6); and (ii) instead of rewriting per-note text in place, it separates immutable Event records from update-first Concept abstractions.

Matched-model comparisons (Section 4.5, Table 5; Appendix Table 16):

  1. LongMemEval-S (N = 500, gpt-4o-mini writer, reader, and official evaluate_qa.py grader): CogniFold reaches 75.4% (377/500), above published 4o-mini scores for Mem0 (53.6%, reproduced by Fang et al., 2025), LightMem (68.6%), and ENGRAM (71.4%), and Zep with a gpt-4o reader (71.2%). Recall and update accuracy stay high with 4o-mini (SSU 95.7%, KU 91.0%, SSA 89.3%), while SSP (66.7% vs. ENGRAM 93.3%), MS (65.4%), and TR (61.7%) account for the gap to our 5.4-mini reader row (93.0%).
  2. Matched write-time controls on LoCoMo and CogEval (Tables 2 and 5): Under shared gpt-4o-mini and embeddings across 3 builds/arm on LoCoMo and 6 configs × 6 scenarios × 3 orders on CogEval (108 builds, 18 per config), Full cuts concept nodes by 42.4% on LoCoMo (and 37.7 → 12.9/scenario on CogEval) vs. add-first Feedback-blocked, lifts Event → Concept Evidence Precision from 0.116 to 0.177 (+0.061, 95% CI [+0.021, +0.102]), and beats Online clustering + naming on LoCoMo (+3.40 points in Table 2; +1.10, CI [+0.04, +2.19] at 4k) and CogEval Harmony (+0.151, CI [+0.085, +0.214]).

2. Contribution of Graph Topology vs. a Mutable Concept-Text Store (W2)

On conversational QA (LoCoMo), read-time edge traversal among concepts contributes very little (-0.32 points); the graph's main QA value comes from bipartite Event to Concept provenance (+10.33 points) and write-time grounding alignment.

We tested both parts of your question directly on our saved builds (Section 4.5, Table 5; Appendix D.2):

LoCoMo Readout Ablation (Same gpt-4o-mini Reader & Judge)J @ 12k (%)J @ 4k (%)Context Tokens / q (4k)
Mutable concept texts only (concepts-only, 3 builds, no Event links)72.9473.041,650
Events-only store (A, raw turns only, no concept layer)81.9580.522,131 (3,475 @ 12k)
Bipartite graph, 1-hop read neighbor expansion disabled (build r1)83.31--4,480 (@ 12k)
Full CogniFold bipartite graph (build r1 / 3-build mean)83.64 / 83.2782.682,515 (-27.6% vs. A@12k)
  1. Inter-concept and read-time topological traversal: Disabling 1-hop graph-neighbor expansion at query time on LoCoMo build r1 moves J from 83.64% to 83.31% (-0.32 points; Multi-Hop 72.70% → 70.21%, and 1,533/1,540 assembled contexts are identical in a deterministic no-decomposition probe). Because hybrid retrieval + reranking already surfaces relevant concepts and events directly, we do not claim that multi-hop edge traversal drives conversational QA. On CogEval's built-in structural QA under a fixed 30%/70% concept/event budget (3,600 chars), restoring typed edges and support-count metadata shows preliminary small-sample gains on state-trajectory questions (3/10 vs. 1/10 without edges) and concept-emergence questions (10/34 vs. 8/34).
  2. Why a flat mutable concept-text store is insufficient: If the store only maintains and modifies concept text without retaining linked Event records (concepts-only), LoCoMo accuracy drops by 10.33 points (83.27% → 72.94%; Temporal -14.6 points, and 69.48% without date/provenance repair), because iterative text updates drop verbatim phrasing and dates. Keeping bipartite Event to Concept links lets Full@4k (82.68%) exceed Events-only@12k (81.95%) on all 3/3 builds with 27.6% fewer context tokens (and Full > A holds across 4o-mini, 4.1-mini, and 5.4-mini judges). Beyond QA, explicit GROUNDS and REINFORCES edges verify 97.1% to 97.5% concept support and supply the degree counts used for candidate-intent filtering (Table 6).

Thank you for the response.After considering it,I’ll keep my original score.

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

This paper introduces CogniFold, an always-on agent memory architecture that organizes incoming events into a three-layer graph consisting of events, concepts, and candidate intents. Its central idea is to continuously consolidate related observations into reusable concepts and generate potential goals from accumulated evidence, rather than relying exclusively on retrieval after a user submits a query. The authors evaluate the approach on LoCoMo and LongMemEval-S for long-term memory question answering and introduce CogEval-Bench to assess concept formation, graph structure, compression, and proactive intent generation.

Critical strengths

The paper addresses a relevant problem for long-running agents: maintaining organized memory and identifying potentially useful actions before explicit requests arrive. The separation between immutable source events and revisable concepts is sensible, as it supports both abstraction and evidence traceability. The update-plan validation and rollback mechanisms also provide a useful foundation for reliable graph operations. Moreover, CogEval-Bench attempts to evaluate memory structure directly rather than relying solely on downstream question-answering accuracy. Under a shared model configuration, CogniFold achieves a Harmony score of 0.476, compared with 0.323 for GraphRAG, suggesting that its organization strategy merits further investigation.

Critical weaknesses
  1. The distinction from existing self-evolving agent methods is insufficiently established. The paper does not clearly explain whether CogniFold represents a fundamentally new self-evolution paradigm or primarily combines incremental memory consolidation with proactive intent generation. The authors should clarify its conceptual and technical differences from existing agent-memory and self-evolving systems, and identify which capabilities cannot be achieved by combining conventional memory with a separate planning or scheduling module.

  2. Reliability under semantic errors is insufficiently evaluated. Atomic execution and snapshot rollback can address certain structural execution failures, but they do not guarantee that LLM-generated concepts and relationships are correct. Incorrect abstractions may influence subsequent memory updates and decisions. The paper lacks systematic experiments on error injection, contradictory observations, erroneous concept merging, cascading failures, and recovery. These are important concerns for a continuously operating memory system.

  3. The evaluation does not fully support the broader claims about proactive cognition and self-evolution. LoCoMo and LongMemEval-S primarily measure memory-assisted question answering. CogEval-Bench provides more relevant structural evidence, but it contains only 251 synthetic events across six scenarios, generated from predefined gold graphs. Furthermore, its Proactivity metric measures the proportion of candidate intents with at least two qualifying relations, not the accuracy or practical utility of those intents. The experiments do not establish that accumulated experience improves subsequent task success, planning quality, or execution efficiency. The LongMemEval-S result also uses a particular model configuration without a matched-reader comparison.

Questions for authors

What specific capability distinguishes CogniFold from a conventional memory system combined with an independent planning module?

How robust is the method when its writer generates incorrect abstractions, and can such errors propagate across subsequent updates?

Can the authors evaluate intent precision, downstream task success, and actual time saved in realistic, long-horizon agent workflows?

2 comments

Thank you for the constructive review and for noting the value of separating immutable events from revisable concepts, atomic rollback, and structural evaluation on CogEval-Bench. Below we address your three concerns with controlled experiments and scope clarifications in Section 4.5 (Tables 5 and 6) and Appendix D.2 (marked in blue).

1. Distinction from Conventional Memory + Separate Planning Module (W1, Q1)

CogniFold performs proactive organization rather than proactive interaction: instead of building an intent-agnostic store that passively awaits queries or bolts on a planner, it continuously tracks user concerns and multi-event support counts inside a bounded graph context, where a separate raw-event planner needs repeated window scans or sees only recent fragments.

We tested both regimes across 18 CogEval runs (6 scenarios × 3 orders) against the 12 reference goals (I*) with the same gpt-4o-mini model and judge (Section 4.5, Table 6 Left):

  • When an independent planner runs unrestricted multi-pass scans over every 10-event raw window (18.3 items/scenario), it gets 80.6% recall, 14.1% precision, and 0.237 Goal F1, outperforming CogniFold's unfiltered per-event proposals (20.3 items, 50.0% recall, 7.4% precision, 0.127 F1).
  • Under a single bounded planning view (k = 8 records, ≤ 1,200 chars), an independent planner over recent raw events drops to 5.6% recall and 8.3% precision (0.065 F1; 0.000 F1 with generic semantic event retrieval), because dispersed supporting events do not co-occur in one small window. Filtering CogniFold concepts to ≥ 3 supporting events (Appendix B.1's 3-event rule) fits recurring themes into the same k = 8 budget, cutting emissions to 1.9 and lifting precision to 23.2% (0.239 F1), at the trade-off of reducing recall from 50.0% to 25.0%.

2. Reliability Under Semantic Errors, Erroneous Merging, and Contradictions (W2, Q2)

Fact-inverting hallucinations are adopted 4× less often than normal concepts (2.0% vs. 8.3%), while forced cross-theme merges continue receiving updates on matching themes (6.3%) without hurting graph Harmony or dual-layer QA.

We ran two stress tests across all 6 CogEval scenarios (Section 4.5, Table 6 Right; Appendix D.2):

  1. Early bad-merge and hallucination injection (t = 25%, 188 subsequent writes): On clean streams, an exposed concept is updated or reinforced on 8.3% of exposures (165/1,995). After ingesting 25% of events, we fused the top-2 distinct concepts into c-err-merged and inserted a fact-inverting concept c-err-halluc (evidence_count=4), then ingested the remaining 75%:
    • Although c-err-halluc and c-err-merged entered the top-15 window on 98.9% and 100.0% of subsequent steps, the writer adopted the contradicting hallucination on only 2.0% of exposures, while c-err-merged was updated on 6.3% (near the 8.3% baseline because incoming events still matched one fused theme).
    • Under the fixed local [email protected] control scorer (Table 3's 0.476 uses [email protected]), final Harmony stayed at 0.363 vs. 0.368 clean, and 1,800-char dual-layer QA across all 65 factual, multi-hop, and state questions held steady: 36/65 (55.4%) injected vs. 35/65 (53.8%) clean (preliminary direct probes on the 6 hallucinated topics: 3/6 dual-layer vs. 0/6 concept-only).
  2. Contradictory observations (12 appended reversals, 24 probes): Appending 2 dated reversals per scenario (e.g., rolling back auth v2.3 to v2.2, stopping ibuprofen), dated UPDATE_NODE + immutable Event links answered 10/12 current-state and 10/12 audit probes correctly, vs. 9/12 and 10/12 for in-place summary overwrite (preliminary small-sample diagnostic).

3. Scope, Multi-Session Intent Validation, and Same-Reader LongMemEval-S (W3, Q3)

We narrowed our claims to memory organization and evaluated both multi-session LoCoMo intent realization and gpt-4o-mini LongMemEval-S.

  • Section 5 notes that we do not measure end-to-end task success or wall-clock time saved in interactive tool-use workflows. Beyond CogEval, we evaluated 176 candidate intents across all 10 LoCoMo conversations (3 gpt-4o-mini builds) against event_summary ground truth using gpt-5.4-mini (90.9% validity agreement with gpt-4.1-mini): 81.8% to 91.1% (mean 85.4%) are valid, and 38.5% to 45.5% (43.1%) are realized 3 to 4.5 sessions later (fresh intents ≤ 2 sessions match a retrieval-only baseline at 40.4% vs. 42.0% at zero query cost; unretired older intents drop to 18.9%).
  • Using gpt-4o-mini as writer, reader, and official evaluate_qa.py grader on LongMemEval-S (N = 500; Table 5 and Table 16) gives 75.4% (377/500), above published 4o-mini results for Mem0 (53.6%), LightMem (68.6%), and ENGRAM (71.4%), and Zep's 4o reader (71.2%). Recall and update categories stay strong with 4o-mini (SSU 95.7%, KU 91.0%, SSA 89.3%), while SSP (66.7% vs. ENGRAM 93.3%), MS (65.4%), and TR (61.7%) reflect reader reasoning differences.

The bounded-window planner comparison clarifies when organizing events into concepts helps, and I appreciate the tests with incorrect merges and contradictory observations. My remaining concern is the practical value of the proposed intents. The LoCoMo analysis shows that some intents anticipate later events. Whether acting on those intents improves planning or task completion remains open, and that distinction matters to my assessment. I maintain my original rating.

Reviewer_9TiJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CogniFold, a graph-based memory for LLM agents with three layers: events, concepts, and candidate intents. Each new event is written into the graph. An LLM writer then updates existing concepts or creates new ones, and may also propose intents linked to supporting evidence. The authors frame this as an “always-on proactive memory”, in which memory continuously self-organizes and intents emerge from the accumulated structure. They evaluate on LoCoMo and LongMemEval-S, and introduce a synthetic benchmark, CogEval-Bench, to measure the structure the memory forms.

Critical strengths

The memory design itself is reasonable. Keeping raw events and abstracted concepts separately, with links between them, makes the memory easy to inspect. The update-first writing reduces redundant concepts while keeping competitive results on LoCoMo and LongMemEval. As a conversational memory system, the results look good.

Critical weaknesses
  1. The core idea is not new, and the paper overclaims it. The main idea this paper emphasizes is that memory should not be static, but should keep updating and reorganizing itself as new information comes in. This has already been explored by memory systems earlier this year, for example EverMemOS, which the paper itself uses as a baseline, and by earlier work before that. Graph-based memory itself is also well explored. Not only has graph-based memory been widely explored, for example HippoRAG, GraphRAG and Zep, which the paper already discusses, but graph-based memory built on abstraction has also been explored. For example, PlugMem abstracts raw experience into knowledge and organizes the memory graph around it. The paper does not discuss this line of work. The paper also describes existing memory systems as having a fixed topology, which I don’t think is a fair description. Once prior work is described accurately, the novelty left is much smaller than the paper suggests.
  2. “Proactive” is used too broadly, and the actual work does not match the claim. In my view, a truly proactive memory should record the right memory at the right time, retrieve the right memory at the right time, and organize the right memory at the right time. That is where proactiveness should show up. In this paper, none of these decisions are really made by the system. Every event is processed the same way, retrieval and consolidation follow a fixed procedure, and the “intents” are simply proposed by the LLM writer during ingestion. The evaluation also does not check whether these intents are correct, useful, or well timed. This is not how proactiveness is usually framed in the proactive-agent literature either, where the key question is whether and when to act. The word “proactive” in the title is much bigger than what the paper actually does.
  3. I don’t understand what “always-on” is supposed to emphasize. This is not a streaming system. Putting this keyword in the title makes readers expect a resident agent working on continuous, possibly multimodal, streaming input. But the experiments are offline replays of text datasets, which is the same setting as previous memory work. Together with the “proactive” framing, I feel the contribution is overstated mainly through the writing, and the writing does not reflect a realistic always-on scenario.
Questions for authors
  1. What exactly does “always-on” mean in this paper, and how is the setting different from the incremental, turn-by-turn ingestion used by prior memory systems? Is there any evaluation under a real streaming setting?
  2. Which part of the system decides when to record, retrieve, or organize memory, or when an intent should be surfaced? CogEval-Bench already defines expected intents, so could the authors report how well the generated intents match them?
  3. How is CogniFold different from prior memory systems that already update and reorganize memory over time, such as EverMemOS and PlugMem? There is also recent work that treats proactive memory as deciding whether to inject a memory-based reminder or stay silent: “Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents” (arXiv, July 2026). I would encourage the authors to discuss how their notion of proactiveness relates to this line of work.
1 comment

Thank you for the direct and substantive review. We agree that our original framing overstated "always-on" and "proactive" and described prior dynamic memory systems too narrowly. We revised Sections 1, 2, 4.4.2, and Appendix B.7 (marked in blue) and ran the evaluations you requested.

1. Prior Dynamic/Abstraction Memories, PlugMem, and Remaining Novelty (W1, Q3)

We removed the "fixed topology" claim, added explicit comparisons with EverMemOS and PlugMem, and isolated write-time folding empirically.

  • EverMemOS (Hu et al., 2026) groups episodic MemCells into MemScenes with Foresight, and PlugMem (Yang et al., 2026) abstracts interactions into linked episodic (G_E), propositional (G_S), and prescriptive (G_P) graphs condensed at query time. Sections 1 and 2 now credit both alongside A-Mem and Zep.
  • CogniFold's specific contribution is write-time self-referential folding under bounded context: rather than extracting window-level propositions and condensing at query time, each incoming batch retrieves existing Concept candidates and applies atomic, rollback-guarded UPDATE_NODE and REINFORCES plans while keeping raw Event nodes immutable.
  • To measure this write-time difference (Section 4.5, Table 5; Appendix D.2), we implemented a matched-gpt-4o-mini version of PlugMem's 6-event window extraction (propositions + prescriptions linked to episodes, omitting its query-time condenser). On CogEval-Bench (~42 events/scenario), window extraction stores 74.0 units/scenario vs. 13.7 folded concepts (-81.5% stored units; preliminary state-tracking QA: 3/10 vs. 2/10). Across 6 configs × 6 scenarios × 3 orders (108 builds, 18 per arm), Full cuts concepts from 37.7 to 12.9 over add-first Feedback-blocked (-42.4% on LoCoMo at 16% lower build cost, USD 2.03 vs. 2.40) and lifts Event → Concept Evidence Precision from 0.116 to 0.177 (+0.061, 95% CI [+0.021, +0.102]).

2. Scope of "Proactive," Relation to Wu et al. (2026), and Matching expected_intents (W2, Q2, Q3)

Our focus is proactive organization rather than proactive interaction, and we evaluated tracked intents directly against expected_intents. Prior memory systems largely build graphs without modeling user intent, passively awaiting queries or bolting an external planning prompt onto raw logs. CogniFold instead builds proactive organization into the graph to continuously track what the user cares about as events arrive, distinct from proactive interaction (e.g., Wu et al., "Remember When It Matters", July 2026), which decides when to interrupt a user or stay silent. Sections 1 and 2 state clearly that our scope is proactive organization rather than proactive interaction, cite Wu et al. (2026), rename Proactivity to Support Incidence in Table 3, and remove the overclaim in Appendix B.7.

To answer Q2, we matched generated intents against CogEval's 12 reference goals (I*) across 18 runs (6 scenarios × 3 orders) and tested degree filtering via Appendix B.1's 3-event rule (Table 6 Left):

Policy (gpt-4o-mini, 18 CogEval runs)Emit / Scen.Recall (%)Precision (%)Goal F1
Ungated per-event writer (Table 3 / Table 12)20.350.07.40.127
Unrestricted periodic planner (every 10-event window)18.380.614.10.237
Bounded view (k = 8, ≤ 1.2k chars): recent raw events1.25.68.30.065
Bounded view (k = 8, ≤ 1.2k chars): concepts with deg ≥ 31.925.023.20.239

Unfiltered per-event proposals over-generate tactical subgoals (20.3 items, 7.4% precision; 47.9% narrower sub-plans), matching your observation. Filtering to concepts with ≥ 3 supporting events cuts proposals by 90% (1.9 items) and raises precision to 23.2% (0.239 F1 vs. 0.065 for k = 8 raw events), at the cost of halving recall (50.0% → 25.0%) and matching rather than beating an unrestricted multi-pass planner (0.237 F1). Across all 10 LoCoMo conversations (3 builds, 176 intents judged by gpt-5.4-mini against event_summary; 90.9% validity agreement with gpt-4.1-mini), 81.8% to 91.1% (mean 85.4%) are valid and 38.5% to 45.5% (43.1%) are realized 3 to 4.5 sessions later. Fresh intents (≤ 2 sessions) match a session-boundary retrieval planner (40.4% vs. 42.0%) at zero query cost, while unretired older intents drop to 18.9%, showing why downstream silence gating (Wu et al., 2026) is needed.

3. Meaning of "Always-On" and Streaming Telemetry (W3, Q1)

We agree our experiments replay text streams rather than resident multimodal daemons. Section 1 now states that always-on refers only to continuous fold-on-arrival without waiting for user queries or global batch rebuilds. Across 251 sequentially ingested CogEval events (Section 4.5), median per-event write latency is 10.0 s with 0/251 rollbacks, write contexts stay bounded (1.9k to 3.6k chars from 25% to 100% prefixes), and mid-stream concept recall reaches 46.1% by the 50% prefix without re-indexing.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.