acceptodds
Under review as a conference paper at ICLR 2027

GRAVITY: Architecture-Agnostic Structured Anchoring for Long-Horizon Conversational Memory

Abstract

Long-horizon memory systems increasingly improve how evidence is stored and retrieved, yet the generator must still reason over fragments whose cross-session relationships are implicit. We study generation-time memory organization as a distinct design dimension and introduce GRAVITY (Generation-time Relational Anchoring Via Injected Topological MemorY), a host-independent auxiliary memory layer. GRAVITY consolidates raw dialogue into entity profiles, temporal event traces, and cross-session topic summaries, then retrieves and injects query-relevant records through the prompt interface. Across five heterogeneous memory systems on LongMemEval and LoCoMo, it improves every host–benchmark baseline under two distinct LLM configurations. Controlled analyses separate gains from organizing already available evidence and from consolidating information across the full history. Under a matched LightMem pipeline, the entity–event–topic representation reaches 83.9% on LoCoMo, 3.6% above the strongest of six alternative auxiliary representations. These results show that generation-time structure is a portable complement to existing memory retrieval, while its interaction with host evidence depends on the benchmark and host.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agent Memory, LLMs, Long Horizon Agents

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_FYRTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GRAVITY, a host-independent auxiliary memory layer that consolidates dialogue into entity profiles, temporal event traces, and cross-session topic summaries. It retrieves these records, expands host queries, and injects selected anchors alongside host memories. The evaluation reports improvements across five memory systems on LongMemEval and LoCoMo under two model configurations, with additional experiments examining evidence organization, source interaction, and alternative auxiliary representations.

Critical strengths

(S1) GRAVITY improves all host–benchmark pairs under both model configurations. Under the primary configuration, average accuracy increases by 9.2% on LongMemEval and 7.5% on LoCoMo. Alternative evaluators also preserve the positive direction of these improvements.

(S2) Anchor construction uses raw dialogue, and the same store is reused across hosts through query and prompt interfaces. This way is practical to augment heterogeneous memory systems without modifying their internal representations.

(S3) The source diagnostics distinguish improvement over host baselines from complementarity with anchors alone. Table 6 shows consistent host contributions beyond anchors on LongMemEval and mixed contributions on LoCoMo. The empirical conclusion is strong.

Critical weaknesses

(W1) That 'same-evidence' baseline doesn't actually prove the evidence is equivalent. 4.3 claims the 80.9% to 82.4% bump is all down to organization, but Appendix A.3.2 points out those records came straight from the retrieved sessions, which may contain answer-bearing facts absent from the compressed host entries. So limiting the sources doesn't cleanly isolate the organization from just finding the evidence.

(W2) The component ablations do not separate complementary views from a larger auxiliary context. The default five-record budget per bank allows up to 5/10/15 anchors in single-/pair-/three-bank configurations. Without matching the total auxiliary-context budget, the full configuration’s advantage does not establish three-view complementarity.

(W3) The uncertainty of the smaller reported advantages is unclear. LoCoMo’s 1540 questions share only ten conversation histories, but no paired uncertainty or conversation-level variation is reported for the 1.5% organization gain or 3.6% representation margin.

(W4) The claimed representation edge seems unproven. While Table 5(a) contrasts different auxiliary setups, it seems to gloss over their extraction schemas, prompt designs, combination rules, retrieval workflows, and context window lengths. Simply keeping the host, model, prompt position, and retrieval budget the same isn't enough to prove the test actually isolates representation.

Questions for authors

(Q1) For Table 5(b), did the extractor receive only the exact memory snippets already shown to the host-only generator, or the complete source sessions containing those snippets? Please clarify the construction inputs and explain why this diagnostic’s host-only accuracy is 80.9%, compared with 77.8% in the main cross-model results.

(Q2) Were total anchor counts and injected tokens matched across the component ablations? Please report these quantities and compare the full configuration against the strongest pair under the same auxiliary-context budget.

(Q3) How consistently do the 1.5% organization gain (Table 5(b)) and 3.6% representation (Table 5(a)) advantage hold across the ten conversations? Please provide conversation-level differences and paired confidence intervals that account for questions sharing the same history.

(Q4) How were the 6 auxiliary representations constructed, consolidated, retrieved, and serialized? Please provide their prompts, schemas, and actual context budgets so that the claimed representation advantage can be assessed independently of differences in extraction and retrieval.

2 comments
AuthorsOpenReview ↗

Response to Reviewer FYRT

Thank you for the very careful review, and for highlighting that GRAVITY improves every host–benchmark pair under both configurations, that one anchor store built from raw dialogue can be reused across hosts in practice, and that the source diagnostics separate gains over the host from complementarity with it. Your questions pointed to several places where the paper was unclear or claimed too much, and we have revised it accordingly.

W1 / Q1. Evidence equivalence and the 80.9% versus 77.8% baselines

Thank you for catching this. The extractor gets the complete retrieved sessions, but so does the baseline: in Table 5(b), the baseline generator also sees the full text of those sessions, not LightMem's compressed entries. We did not say this clearly.

Concretely, let S be the full text of the sessions behind LightMem's retrieved entries. The baseline sees S, the records R(S) are built from S only, and the augmented model sees S plus R(S), with query expansion turned off. Both sides have access to exactly the same dialogue, so the 80.9% to 82.4% gain shows that adding records helps without widening the source evidence. This setup rules out evidence from outside S, but it does not separate the structure of the records from the extra processing, repeated presentation, and longer input they also bring.

Generation inputAccuracy
S80.9%
R(S)77.1%
S + R(S)82.4%
R(F), F = full history80.3%
S + R(F)85.5%

This also answers why the baseline is 80.9% here but 77.8% in Tables 2 and 5(a): those tables give the generator LightMem's compressed entries, while this one gives it the full session text. For the same reason, S + R(F) (85.5%) is higher than LightMem + GRAVITY in the main results (83.9%): the former pairs the records with full session text, the latter with compressed memories, so the two are different input settings (Appendix A.3.2). In the revision, the rows of Table 5(b) are labeled by what the generator sees, and the protocol is spelled out in the caption (L379–383), Section 4.3 (L353–361), and Appendix A.3.2 (L917–940).

W2 / Q2. Total context budget in the component ablations

No, they were not matched, and you are right that this matters. With up to five records per bank, configurations with more banks simply received more context:

Original ablationEVTTotal recordsAuxiliary tokens
E4.30––4.30153.4
V–4.82–4.82381.0
T––4.714.711185.0
EV4.304.82–9.12534.5
ET4.30–4.719.001338.5
VT–4.824.719.521566.0
EVT4.304.824.7113.821719.5

We therefore reran the comparison between all three banks (EVT) and the two best pairs (ET and VT, both 74.0% in Table 3) under the same per-question input cap, on all 1,540 LoCoMo questions with LightMem. The host memories were identical across conditions (query expansion off, so EVT here matches the −exp row of Table 3). For each question, EVT's prompt length set the cap, and ET and VT filled the freed space with more records from their own banks.

ConditionAccuracyRecords by bankAuxiliary tokensBudget used
EVT75.26%E 4.30; V 4.82; T 4.711,719.5100%
ET74.61%E 7.21; T 5.551,639.294.0%
VT74.48%V 6.26; T 4.801,686.597.8%

EVT has a slightly higher observed accuracy, by 0.65 and 0.78 points, but we report no paired intervals for these differences, and ET and VT used somewhat less of the auxiliary budget. We therefore read the three configurations as close: part of the original gap came from the extra context, and the remaining difference is not evidence of three-view complementarity. We have added these results to Appendix A.8.4 (L1751–1798) and softened the claim in Section 4.3 (L316–323) and Appendix A.3.3 (L963–965).

AuthorsOpenReview ↗

Response to Reviewer FYRT (cont.)

W3 / Q3. Uncertainty across conversations

We now report conversation-level results for both margins. We resampled the ten conversations 10,000 times, keeping each conversation's questions together and pairing the two conditions:

ComparisonAccuracyDifference95% CIConversations with a gainLeave-one-out range
S + R(S) vs. S80.91% → 82.40%+1.49[0.52, 2.40]8 of 101.27 to 1.85
GRAVITY vs. GSW-style80.32% → 83.90%+3.57[2.90, 4.32]10 of 103.31 to 3.78

Per-conversation differences (percentage points):

Conversation26304142434447484950
S + R(S) − S+1.32−1.23+1.32+3.02+0.56+4.07+2.67−1.05+1.92+1.90
GRAVITY − GSW-style+4.61+3.70+1.97+3.52+3.37+3.25+6.00+2.09+4.49+3.16

Neither interval includes zero, and dropping any single conversation does not change the sign. The smaller 1.5-point gain is less consistent, with two conversations going the other way, while the 3.6-point margin holds in every conversation. Section 4.3 now cites both intervals (L346–361), and full counts are in Appendix A.8.5 (L1800–1835).

W4 / Q4. Construction and comparability of the auxiliary representations

We agree the original description was too thin. Table 31 (Appendix A.8.6, L1836–1941) now describes, for every pipeline, how it is built and merged, what text it can see, how it retrieves, what the answer model actually receives (including caps and truncation), and how many auxiliary tokens it adds. The prompts are in the anonymous code repository linked in Appendix A.1 (L762), under auxiliary_pipelines/prompts/, together with the shared answer and judge templates. GraphRAG- and LightRAG-style use the default prompts of langchain-graphrag 0.0.9 and lightrag-hku 1.5.4, and GSW-style uses the prompts from its public code.

All pipelines share the LightMem store, top-60 retrieval, answer template, and generator (Claude Sonnet 4.6), with GRAVITY in its default configuration. A short version:

PipelineBuilt fromWhat the answer model seesAux. tokens
Free-form summaryRetrieved sessionsA free-form summary of those sessions1,205
GRAVITYFull historyUp to 5 entity, 5 event, and 5 topic records2,673
GraphRAG-styleFull history15 entities, 20 relations, 5 community reports3,043
RAPTOR-styleFull historyUp to 5 tree nodes2,314
HippoRAG2-styleFull history5 facts and 8 passages1,807
LightRAG-styleFull historyUp to 30 entities, 30 relations, 15 chunks5,526
GSW-styleFull historyUp to 15 actors, states, and operators; 10 QA pairs; background2,156

(Aux. tokens: mean per question on conv-26, 152 questions, under the cross-model configuration of Table 5(a). The 1,719.5 tokens for EVT under W2/Q2 are measured under the GPT-4o-mini configuration on all LoCoMo questions, so the two numbers are not directly comparable.)

As the table shows, the pipelines differ in more than representation: in what they are built from, how they retrieve, and how much they put in the prompt. You are right that our setup does not isolate representation. We also think it cannot be fully isolated, since each representation comes with its own way of being built and searched. We have therefore narrowed the claim. Under a shared host, template, and generator, GRAVITY outperforms each of these pipelines, including the GSW-style one that reads the complete history (+3.57 points, positive in all ten conversations), but we no longer say this isolates representation alone. The wording is updated in the abstract (L23–25), introduction (L63–65, L77–78), Related Work (L99–101), Section 4.3 (L346–352), Discussion (L517–519), Conclusion (L534–537), and Appendix A.3.2 (L949–952).

Reviewer_i7U7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GRAVITY (Generation-time Relational Anchoring Via Injected Topological MemorY), a host-independent auxiliary memory layer for long-horizon conversational memory. GRAVITY consolidates raw dialogue into entity profiles, temporal event traces, and cross-session topic summaries, retrieves query-relevant anchors, and augments existing memory systems through query expansion and prompt-level context injection without modifying their internal memory architectures. Experiments across five heterogeneous memory systems, LongMemEval and LoCoMo, and two model configurations report consistent improvements over host baselines. Component ablations, alternative auxiliary representation comparisons, and evidence-source controls further investigate the sources of these gains.

Critical strengths
  1. GRAVITY provides an auxiliary memory layer independent of the host system; this layer is integrated via query expansion or context injection at the prompt level, without requiring modifications to the host system's internal memory architecture.

  2. The evaluation encompassed five memory systems with distinct characteristics, two benchmarks, and two model configurations. By conducting component ablation (i.e., removing or disabling specific components) and controlling the sources of evidence, the study gained valuable insights into performance improvements.

Critical weaknesses
  1. Lack of direct comparison with similar pluggable memory enhancement methods. The paper mainly compares the performance of various systems before and after adding GRAVITY. Existing works such as MemR³ [1] and MemMA [2] also enhance existing memory systems through additional modules. Although the current results demonstrate the effectiveness of GRAVITY, its advantages and improvements compared to similar works are unclear. It is recommended to supplement experiments with similar comparative methods and unify the model, evidence scope, and computational budget.

[1] MemR3: Memory Retrieval via Reflective Reasoning for LLM Agents [2] MemMA: Coordinating the Memory Cycle through Multi-Agent Reasoning and In-Situ Self-Evolution

  1. The source of the performance gains due to the additional token budget is unclear. Table 18 in Appendix A.3.6 shows that under the main GPT-4o-mini configuration, the average total token consumption per query for LightMem on LoCoMo increased from 2,313 to 4,211, an increase of approximately 82%. Since adding GRAVITY leads to such an increase in tokens, the baseline tokens should also be increased by the same amount in the comparison. Otherwise, there may be more cost budget leading to performance improvements, but the reason for this is unclear.
Questions for authors

The paper employs the "LLM-as-judge" approach as the primary evaluation method and assesses the robustness of the results by having different models score the same set of outputs. While this mitigates reliance on a single judge, it does not verify the alignment between automated scoring and human judgment. Based on my understanding and experience, methods utilizing LLM-as-judge typically incorporate human verification; therefore, I recommend adding a human evaluation component.

5 comments
AuthorsOpenReview ↗

Response to Reviewer i7U7

Thank you for recognizing that GRAVITY works as an independent layer that plugs into existing memory systems without changing them, and that the evaluation across five hosts, two benchmarks, and two model configurations, together with the ablations and source controls, gives useful insight into where the gains come from. Your suggestions led to three new experiments, described below.

W1. Comparison with MemR³ and MemMA

Thank you for these references; we agree they are the right comparisons. MemR³ improves retrieval by having the model reflect on what evidence is still missing and search again. MemMA coordinates how memory is built and searched, and checks and repairs the memory using probe questions. GRAVITY instead adds a separate cross-session representation that is reused at answer time. We have added both papers to Related Work (L103–124).

We ran both on all 1,540 LoCoMo questions, using the same LightMem backend, the same GPT-4o-mini generator and judge, and the same conversation histories. Both methods are run from their official code, with all LLM roles using GPT-4o-mini. For MemMA, we use the official LightMem self-refine pipeline, which keeps construction guidance, probe-based verification and repair, and iterative query refinement. To use one base model throughout, its probes are generated online by GPT-4o-mini, five per session, instead of the Claude Opus 4.5 probes of the original paper. At query time, the Meta-Thinker judges answerability at most H = 3 times (the repository's example scripts use 5 on a 10% subset) and stops earlier once it judges the question answerable or cannot form a new query. For MemR³, the official retrieve, reflect, and answer controller with evidence-gap tracking is connected to LightMem through a retrieval adapter. Each retrieval returns 60 new entries, the controller is called at most five times with at most one consecutive reflection, and it stops once it judges the evidence sufficient or reaches this limit. To keep the comparison fair, the final answer prompt of MemMA and MemR³ was capped at the length of GRAVITY's prompt for the same question. Full configurations and run statistics are in Appendix A.8.3 (L1690–1712).

MethodAccuracy (%)Gain over LightMemOnline LLM tokens/queryOffline tokens/conversationTotal tokens/query
LightMem70.13–2,313–2,313
+ MemMA71.36+1.236,365979K12,722
+ MemR³73.38+3.257,10207,102
+ GRAVITY75.78+5.654,211193K5,462

Online tokens cover all LLM calls made to answer a question, including intermediate reasoning calls and excluding judging; for GRAVITY this is the final generation. Offline tokens exclude LightMem's own memory construction, which all methods share. Total tokens per query add the offline tokens spread over the questions about each conversation.

GRAVITY beats MemMA by 4.4 points and MemR³ by 2.4 points. Both differences hold up when we resample conversations (95% CIs [2.41, 6.68] and [0.80, 4.10]), and GRAVITY wins in 9 and 8 of the 10 conversations. GRAVITY also uses the fewest tokens of the three enhancements. Online, it makes one LLM call per question, while MemMA and MemR³ make several. Offline, MemMA uses about five times as many tokens as GRAVITY, and MemR³ needs no construction beyond LightMem. GRAVITY's 193K construction tokens are paid once per conversation, which adds about 1.25K tokens per LoCoMo question; with this included, GRAVITY uses 5,462 tokens per question, compared with 7,102 for MemR³ and 12,722 for MemMA. The full comparison is in Section 4.3 (L367–373) and Appendix A.8.3 (L1671–1748).

W2. Additional token budget

This is a fair concern, so we tested it directly on LightMem and LoCoMo. For each question, we took the exact number of input tokens GRAVITY used and gave LightMem that same budget, filled with more of its own retrieved memories (122 instead of 60, at the median). Everything else stayed the same, and no anchors were added. The budget was almost fully used (99.6% on average).

ConditionAccuracy (%)Tokens/query
LightMem70.12,313
LightMem + more retrieved memories71.84,208
LightMem + GRAVITY75.84,211

With the same generation budget, more memories add 1.7 points and GRAVITY adds 5.7, so the extra tokens alone do not explain the gain. This is now in Section 4.3 (L367–373) and Appendix A.8.2 (L1654–1669).

AuthorsOpenReview ↗

Response to Reviewer i7U7 (cont.)

Q. Human verification of LLM-as-judge evaluation

Our primary evaluation follows the LLM-as-judge protocol of LightMem, so that results stay comparable with the hosts' reported numbers. We agree that human verification is needed and have added a human evaluation (Section 4.5, L425–428; Appendix A.8.1, L1629–1651). We sampled 100 questions from each benchmark and took the existing answers from LightMem and from LightMem + GRAVITY, for 400 answers in total. Three annotators who are not authors of the paper judged each answer against the reference and the dialogue, without knowing which system produced it or what the LLM judge said, and we took the majority vote. It agrees with the GPT-4o-mini judge on 97% of the answers. The judge also does not favor GRAVITY: on the same sample, the gain under human labels is at least as large as under the judge.

BenchmarkLabelsLightMemLightMem + GRAVITYGain
LoCoMoLLM judge70%75%+5
LoCoMoHuman69%76%+7
LongMemEvalLLM judge71%73%+2
LongMemEvalHuman68%72%+4

(100 answers per method and benchmark.)

AuthorsOpenReview ↗

Kindly reminder

Dear Reviewer i7U7,

Thank you again for your review. As the discussion period is coming to a close, we wanted to check whether our responses have addressed your concerns. In brief, we have added:

  • A direct comparison with MemR³ and MemMA using the same LightMem backend, models, and final-input cap. GRAVITY scores 75.78%, compared with 73.38% and 71.36%, with conversation-level paired 95% CIs excluding zero, and uses the fewest total tokens per question of the three (Appendix A.8.3).
  • A matched-budget control. Giving LightMem GRAVITY's per-question token budget filled with more retrieved memories reaches 71.8%, compared with 75.8% for GRAVITY (Appendix A.8.2).
  • A blinded human evaluation of 400 answers. Human labels agree with the LLM judge on 97% of them, and the gain under human labels is at least as large as under the judge (Appendix A.8.1). If any of these points would benefit from further clarification, or if there is anything else we can provide, we would be glad to do so before the discussion period ends.

Best regards,

The authors

Reviewer_t6RDOpenReview ↗
Rating
3
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GRAVITY, an architecture-agnostic auxiliary memory layer that helps models connect retrieved information across conversations. It organizes dialogue into entity profiles, event trajectories, and topic summaries, integrating them with existing memory through query expansion and prompt augmentation without modifying the underlying architecture. Across five memory systems on LongMemEval and LoCoMo, GRAVITY improves overall accuracy under both model configurations, with average gains of 9.2 and 7.5 percentage points, respectively, in the main configuration. Controlled experiments attribute these gains to both structured organization of existing evidence and cross-session integration of the full conversation history, while the proposed representation outperforms six alternatives under a common evaluation protocol. Its complementarity with existing memory varies by task and system, and LLM-based extraction and summarization may lose details, distort chronology, or conflate entities.

Critical strengths

The paper is well structured and presents extensive experimental results showing substantial performance improvements.

Critical weaknesses

Unclear motivation and computational overhead. The motivation for introducing a separate GRAVITY layer is not fully clear to me. More importantly, the paper does not clearly report the additional token usage or runtime required to construct and use its entity and event anchors. Without this information, it is difficult to assess whether the performance gains justify the added cost.

Limited model coverage. The main experiments rely on the relatively old GPT-4o mini. It is unclear whether the observed gains would hold with newer, stronger models or whether these models would exhibit different patterns.

Questions for authors

How do you ensure fair comparisons between GRAVITY-augmented systems and their baselines, particularly regarding model configurations and token or compute budgets?

What additional token usage and runtime does GRAVITY introduce, both during anchor construction and at inference time? How do the performance gains compare under matched resource budgets?

3 comments
AuthorsOpenReview ↗

Response to Reviewer t6RD

Thank you for your positive assessment, and for noting that the paper is well structured and that the experiments show substantial improvements. Your questions on motivation, cost, and fairness helped us make these aspects much clearer in the revision.

W1 / Q2. Motivation and computational cost

Why a separate layer. Memory systems have become good at retrieving relevant fragments, but the generator still has to work out how those fragments relate: which facts belong to the same person, what changed over time, and which sessions are about the same thing. GRAVITY builds these links ahead of time as entity profiles, event traces, and topic summaries. Because it only uses raw dialogue and the host's query and prompt interfaces, the same store can be built once and attached to any host. In our experiments it improves all five hosts on both benchmarks.

Three results in the paper suggest that retrieving more fragments is not enough by itself. First, the oracle diagnostic shows that anchors help most when the relevant evidence is mixed with distractors (Table 19). When the prompt contains all gold evidence, scattering it among distractors drops accuracy from 84.9% to 75.6%, and adding anchors recovers 2.9 points. With the gold evidence alone, however, anchors slightly lower accuracy (84.9% to 83.9%), so the benefit lies in reducing the effect of distracting context, not in adding structure to clean evidence. Second, when the baseline and the extractor read the same retrieved sessions, adding records improves accuracy from 80.9% to 82.4% (Table 5(b)). Third, spending GRAVITY's extra tokens on more retrieved memories instead brings LightMem only to 71.8%, compared with 75.8% with GRAVITY (see Q1/Q3 below).

Cost. We do report these numbers, but they are in the appendix and easy to miss: Table 16 has construction cost, Table 18 has inference cost, and Table 7(b) has cheaper build variants. We agree they belong in the main text, and the revision now summarizes them in a new Table 7(c) (L448–451) and in Section 4.5 (L430–466).

Anchors are built once per conversation, offline (GPT-4o-mini, all three banks):

BenchmarkInput tokensOutput tokensTotal tokensTimeAPI cost
LoCoMo152.4K40.2K192.6K556.8 s≈ $0.047
LongMemEval90.6K27.2K117.8K360.6 s≈ $0.030

The same store is then reused by every host and every question about that conversation. For comparison, the hosts' own memory construction takes 114K to 1,693K tokens per LoCoMo conversation, so GRAVITY is cheaper than four of the five and faster than all of them.

At inference time, GRAVITY makes no extra LLM calls, since anchor selection and query expansion use only matching and embeddings. Its cost is a longer prompt. For LightMem, tokens per query go from 2,313 to 4,211 on LoCoMo and from 819 to 1,223 on LongMemEval, and latency goes up by about half a second. Across hosts, the added latency is under one second except for ZEP on both benchmarks and LiCoMemory on LongMemEval, whose graph-based retrieval spends extra time running the expanded queries. If build cost matters, running the extractors in parallel halves build time with no accuracy change, and a fused extractor cuts build tokens by 75% at a cost of 2.7 points (Table 7(b)).

Is it worth it? For LightMem on LoCoMo, roughly $0.05 of one-time construction per conversation and about 1.9K extra tokens per query take accuracy from 70.1% to 75.8%. Spending the same tokens on more retrieved memories only gets to 71.8%.

Q1 / Q3. Model and resource controls

Within each configuration, a host and its GRAVITY version use the same generator, judge, answer template, and retrieval limit (60 memories on LoCoMo and 20 on LongMemEval, as in LightMem). Query expansion swaps memories within that limit rather than adding more, and anchors go into the prompt as extra context.

Since the extra context means extra tokens, we added a matched-budget control for LightMem on LoCoMo. For each question, we take the number of input tokens GRAVITY used and fill that budget with more of LightMem's own retrieved memories instead of anchors (on average 99.6% of the budget is used). This gets 71.8% at 4,208 tokens per query, while GRAVITY gets 75.8% at 4,211, so GRAVITY stays 4 points ahead at nearly identical generation-token usage (Section 4.3, L367–373; Appendix A.8.2, L1654–1669). This control matches the input budget of final answer generation only; GRAVITY additionally requires the one-time anchor construction reported under W1/Q2 above (192.6K tokens per LoCoMo conversation).

AuthorsOpenReview ↗

Response to Reviewer t6RD (cont.)

W2. Evaluation with a stronger model configuration

We did test a stronger model. Table 2 reruns all five hosts on both benchmarks with Claude Sonnet 4.6 for generation and extraction and GPT-5.6 sol as the judge. GRAVITY helps in all ten cases, by 15.2 points on average on LongMemEval and 18.1 on LoCoMo.

A few patterns do change with the stronger model. The gains get larger (versus 9.2 and 7.5 under GPT-4o-mini), though the judge also changed, so we would not read too much into the size difference. Weaker hosts benefit more (Figure 2). And the anchor store on its own becomes quite strong, at 80.9% on LoCoMo and 79.0% on LongMemEval (Table 6). This configuration was easy to overlook, so Section 4.1 now says which tables use which models (L260–261), and Section 4.2 discusses these patterns (L300–301).

AuthorsOpenReview ↗

Kindly Reminder

Dear Reviewer t6RD,

Thank you again for your positive assessment. As the discussion period is coming to a close, we wanted to check whether our responses have addressed your questions. In brief:

  • Cost. Anchor construction is done once per conversation and takes about 193K tokens and 557 seconds on LoCoMo (about $0.05), fewer tokens than four of the five hosts' own memory construction. At inference, GRAVITY makes no extra LLM call and adds about 1.9K prompt tokens per LoCoMo query. These costs are now summarized in the main text (Table 7(c), Section 4.5).
  • Matched budget. Giving LightMem GRAVITY's per-question generation-token budget, filled with more retrieved memories, reaches 71.8%, compared with 75.8% for GRAVITY (Appendix A.8.2).
  • Stronger model. Table 2 evaluates all five hosts on both benchmarks with Claude Sonnet 4.6 and GPT-5.6 sol, and GRAVITY improves all ten host–benchmark pairs; Section 4.1 now states which tables use each configuration.

If any of these points would benefit from further clarification, we would be glad to provide it before the discussion period ends.

Best regards,

The authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all three reviewers for their careful and constructive reviews. We are glad that the reviewers found GRAVITY's improvements consistent across all hosts, benchmarks, and model configurations (t6RD, FYRT), its design practical because one anchor store can be attached to different memory systems without changing them (i7U7, FYRT), and the evaluation extensive, with ablations and source controls that give useful insight (t6RD, i7U7, FYRT).

We have uploaded a revised PDF in which all new or rewritten text is in blue. Most new results are in a new Appendix A.8 (Tables 23–31), with short summaries in Sections 4.3 and 4.5. We kept all existing table, figure, and section numbers, so references to the original submission still work. Line numbers below refer to the revised PDF.

ChangeLocation in the revised PDF
Narrowed claims on Table 5(a)Abstract (L23–25); Introduction (L63–65, L77–78); Related Work (L99–101); Section 3.1 (L163–164); Section 4.3 (L346–352); Table 5 (L384); Discussion (L517–519); Conclusion (L534–537)
MemR³ and MemMARelated Work (L103–124); Section 4.3 (L367–373); Appendix A.8.3 (L1671–1748, Tables 25–26)
Matched generation budgetSection 4.3 (L367–373); Appendix A.8.2 (L1654–1669, Table 24)
Human evaluationSection 4.1 (L249–250); Section 4.5 (L425–428); Discussion (L515–516); Appendix A.8.1 (L1629–1651, Table 23)
Cost in the main textSection 4.5 (L430–466); new Table 7(c) (L448–451)
Model configurationsSection 4.1 (L260–261); Section 4.2 (L300–301)
Table 5(b) protocol and wordingAbstract (L21–23); Introduction (L44–46, L71–73); Table 5 caption (L379–383); Section 4.3 (L353–361); Appendix A.3.2 (L917–940); Conclusion (L534–537)
Budget-matched ablationSection 4.3 (L316–323); Appendix A.3.3 (L963–965); Appendix A.8.4 (L1751–1798, Tables 27–28)
Conversation-level uncertaintySection 4.3 (L346–361); Appendix A.8.5 (L1800–1835, Tables 29–30)
Pipeline details and contextAppendix A.8.6 (L1836–1941, Table 31)

In short, we added:

  1. A matched-budget control: giving LightMem the same generation-token budget as GRAVITY, but filled with more retrieved memories, gets 71.8%, compared with 75.8% for GRAVITY (t6RD, i7U7).
  2. A head-to-head comparison with MemR³ and MemMA on the same backend and models. GRAVITY scores 75.78%, against 73.38% and 71.36% (i7U7).
  3. A human evaluation: the majority vote of three non-author annotators agrees with the LLM judge on 97% of 400 sampled answers (i7U7).
  4. A clarification of Table 5(b): the baseline and the record extractor read exactly the same session text (FYRT).
  5. A component ablation under the same per-question input cap. The three configurations come out close (75.26%, 74.61%, and 74.48%), so we no longer treat the remaining gap as evidence of three-view complementarity and have toned down that claim (FYRT).
  6. Conversation-level confidence intervals for the 1.5- and 3.6-point margins (FYRT).
  7. Full implementation details for all auxiliary pipelines, with their prompts in the anonymous code repository linked in Appendix A.1 (L762), under auxiliary_pipelines/prompts/ (FYRT).

Thank you for the follow-up, and for confirming that the paired intervals address the uncertainty concern. We agree with points 2 and 4, and the revision already reflects them: we no longer claim three-view complementarity, and we no longer describe Table 5(a) as isolating representation. We would like to clarify how the remaining points relate to the paper's main claim. The main claim is that a single anchor store, built once from raw dialogue, improves heterogeneous memory systems without changing them. This rests on positive gains in all ten host–benchmark pairs under two model configurations and three evaluators, as you noted in S1 and S2. The diagnostics in Section 4.3 examine where these gains come from, and the revised text states their scope explicitly.

On point 1, we ran an additional control that targets these factors directly. The generator receives S together with a free-form LLM summary of the same sessions. For each question, the summary is generated with the token count of the corresponding R(S) as its target length and then truncated to that length. This keeps the source evidence and the input length the same as S + R(S) and also adds LLM-processed text, but without the entity, event, and topic structure. It reaches 81.0% (1,247/1,540), essentially unchanged from S alone (80.9%, 1,246/1,540), whereas S + R(S) reaches 82.4% (1,269/1,540). With the same conversation-level paired bootstrap as before, S + R(S) exceeds S + summary by 1.43 points (95% CI [0.58, 2.24]); it is higher in eight conversations, tied in one, and lower in one, and leaving out any single conversation gives 1.20 to 1.78 points. Adding LLM-processed text of the same length therefore leaves accuracy unchanged, while adding the structured records accounts for almost all of the 1.5-point gain over S. This is consistent with the matched-budget control in Appendix A.8.2, where spending GRAVITY's per-question token budget on more retrieved memories raises LightMem only from 70.1% to 71.8%, versus 75.8% with GRAVITY. We have added this control to the revision as Appendix A.8.7.

On point 2, we agree that the budget-matched ablation does not establish three-view complementarity, and the revision no longer claims it. The main results do not rely on this claim. We keep all three banks as the default because this configuration is at least as accurate as the best pair in every setting we evaluated (both benchmarks and both model configurations; Tables 3 and 4) and does not lose accuracy under a matched budget (Appendix A.8.4). We therefore present it as a robust default rather than as evidence that each view is necessary.

On point 4, we agree that the comparison is between auxiliary pipelines. Each representation comes with its own extraction and retrieval procedure, so these cannot be held fixed while the representation changes. We therefore state the result as a pipeline-level comparison under a shared host, answer template, and generator, which is how it now appears in the abstract, Section 4.3, and the Discussion.

Thank you again for the careful review, which has made the claims in the paper more precise. We would welcome further discussion, and if there is a specific analysis or control that would help resolve the remaining concern, we would be grateful for the suggestion and glad to run it within the discussion period.

Dear Reviewers,

As the author–reviewer discussion closes today, we thank you again for your time and feedback. For convenience, the main additions are summarized in our General Response above, with line numbers in the revised PDF:

  • Comparison with MemR³ and MemMA under the same backend, models, and final-input cap: GRAVITY is higher by 2.4 and 4.4 points and uses the fewest total tokens (Appendix A.8.3).
  • Matched generation budget: 71.8% for LightMem with more retrieved memories, versus 75.8% for GRAVITY (Appendix A.8.2).
  • Length-matched unstructured control: 81.0% versus 82.4% for the structured records, +1.43 points with 95% CI [0.58, 2.24] (Appendix A.8.7).
  • Blinded human evaluation: 97% agreement with the LLM judge (Appendix A.8.1).
  • Cost summary in the main text (Table 7(c)) and narrowed claims on three-view complementarity and representation isolation.

We would be grateful if you could take these into account in your final assessment. Thank you.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.