acceptodds
Under review as a conference paper at ICLR 2027

Retrieval-Driven Memory Reconsolidation for Long-Term LLM Agents

Abstract

Long-term memory is essential for LLM-based agents operating over extended interactions. Many existing systems primarily update memory when new information arrives, treating retrieval as the endpoint of memory access rather than a driver of memory evolution. Recent methods have begun to exploit signals from memory use, but it remains less explored how retrieved evidence can guide agent-selected edits to a typed memory structure. Inspired by memory reconsolidation in cognitive neuroscience, we propose REALM, a reconsolidation-evolution agentic long-term memory framework. REALM organizes experiences into a heterogeneous cognitive graph, retrieves evidence through adaptively composed graph-search atoms, and performs retrieval-driven reconsolidation by using the activated subgraph and interaction feedback to add, strengthen, or weaken typed relations. Under our evaluation protocol, REALM achieves average accuracies of 75.97% on LoCoMo and 65.11% on LongMemEval, corresponding to gains of 7.17 and 1.90 points over the strongest reported baseline averages. Ablation studies show that memory reconsolidation improves overall performance on both benchmarks, while further analyses indicate that it progressively reorganizes related memory units into more coherent local structures for collective evidence recall and utilization during reasoning. These results support retrieval-driven reconsolidation as one mechanism for adaptive long-term memory in LLM agents.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM-based Agents, Long-Term Memory, Memory Reconsolidation, Continual Learning

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_29DqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes REALM, a retrieval-driven memory system with a cognitive graph, adaptive search, and post-answer reconsolidation. The core idea is to update the activated subgraph after each QA step so memory evolves with use. Experiments on LoCoMo and LongMemEval report gains from reconsolidation, but also clear failures on preference-style questions.

Critical strengths
  1. The overall design is fairly complete, and the closed-loop setup matches a reasonable cognitive story. Claims of automatic organization, adaptive retrieval, and memory enhancement also line up with practical needs.
  2. The method works with multiple memory organizations and retrieval styles, which makes the system more usable in real engineering settings.
Critical weaknesses
  1. Reconsolidation feedback includes the ground-truth answer text, not only a correct/incorrect bit. The topology update is therefore driven by oracle labels. That is much stronger than the “task feedback in practice”, and it makes the closed-loop claim hard to trust in realc system.
  2. The gains are not especially large, and they are uneven across benchmarks. LoCoMo improves by about 7 points over the strongest reported average, but LongMemEval only by about 2, and reconsolidation itself adds roughly 2 points. Preference is far behind strong baselines. So the method does not look generally better, it looks stronger on some settings and weak on others.
Questions for authors
  1. With feedback reduced to correct/incorrect only, and without the gold answer string, how do Tables 1–3 change?
  2. Appendix C says preference fails mainly because seed retrieval misses the topic, so reconsolidation cannot help if the subgraph is never activated. If the gold preference evidence is forced into the seed set, how high does preference accuracy go?
Reviewer_81ySOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Most LLM-agent memory systems change memory only when new information arrives. Retrieval reads memory but leaves it unchanged. Taking inspiration from reconsolidation in neuroscience, the paper proposes REALM, which also edits memory after it has been used. REALM has three parts: Organization, Retrieval and Reconsolidation. The main contribution of the paper revolves majorly around "Reconsolidation" phase, in which after each answer, an LLM sees the retrieved subgraph, and q,a^ and a (question, generated answer and golden answer). It then adds, strengthens or weakens edges among the retrieved nodes. The method evaluates on two memory benchmarks i.e LoCoMo and Longmemeval (s) showing gains over some existing agent memory systems.

Critical strengths

S1. A clear, inspectable reconsolidation operator. The edit step is well defined. After each question, an LLM may add, strengthen or weaken typed, weighted edges among the retrieved nodes. Updates are bounded (Eqs. 11–12), and nodes are never edited and edges never deleted (Prompt F.7). Keeping edits local and leaving stored content untouched is a sensible safety choice. Because the agent chooses typed edits, the operator also differs from simple co-activation rules, and it could be added to other graph memories.

S2. Targeted analyses and candid reporting. The paper isolates reconsolidation with an on/off ablation on both benchmarks (Table 3) and tests sensitivity to question order (Fig. 3). It also analyzes how relation types and expansion depth are used during retrieval (Figs. 4–5). Appendix C openly discusses failure modes, including harmful updates and retrieval misses on preference questions. Most prompts and hyperparameters are included.

Critical weaknesses

W1. The reconsolidation gain relies on gold labels and oracle sessions, and it is small. Reconsolidation receives each question's reference answer and correctness label (Sec 2.4.1), which no baseline gets. On LoCoMo, questions are answered in sequence, and about half share gold evidence with an earlier question. Labels from earlier test questions can therefore steer later retrieval. On LongMemEval, the warm-up questions are generated only from the oracle sessions, which contain the evidence (App. B.4.3). The generator prompt also uses six evaluated test questions, with their answers, as examples (Prompt B.1). The gain is also small. Tables 1 and 3 match the most favourable of the four question orders in Fig. 3; averaged over the four orders, the LoCoMo gain is about +1.1, not +2.01. On LongMemEval it is 10 of 470 questions (296→306), with no variance reported. Results without gold feedback, with warm-ups from sessions that don't contain the evidence, and over multiple runs are needed to support the claim

W2. The baselines are not compared like-for-like, and stronger results are left out. Sec 3.1 claims "identical judge prompts and models", but most baseline numbers are copied from other papers (App. B.5), several of which use different judges. On LoCoMo, MAGMA (the basis of the +7.17 claim) and Nemori are scored with MAGMA's graded 0–1 rubric, not REALM's binary MemOS judge. MIRIX and Mem0 come from MemOS's Table 3 [4], which also reports MemOS itself at 75.80, a margin of only 0.17. On LongMemEval, all baselines are copied. Where this can be checked, their sources use the full 500 questions rather than REALM's 470, and Zep was judged by GPT-4o. Averages are recomputed for some rows but not others, and stronger systems in MemOS's Table 4 (77.8, 72.4, 66.4) are left out. Even as printed, REALM leads in only two of six categories. The related work also omits HiMem [1] and [5], which already reconsolidate memory after retrieval, and Memory-R1 [2] and ReasoningBank [3], which use task outcomes to shape memory.


References [1] N. Zhang, X. Yang, Z. Tan, W. Deng, and W. Wang. HiMem: Hierarchical Long-Term Memory for LLM Long-Horizon Agents. arXiv preprint arXiv:2601.06377, 2026.

[2] S. Yan, X. Yang, Z. Huang, E. Nie, Z. Ding, et al. Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning. arXiv preprint arXiv:2508.19828, 2025.

[3] S. Ouyang, J. Yan, I-H. Hsu, Y. Chen, K. Jiang, et al. ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory. In International Conference on Learning Representations (ICLR), 2026. arXiv:2509.25140.

[4] Z. Li, C. Xi, C. Li, D. Chen, B. Chen, et al. MemOS: A Memory OS for AI System. arXiv preprint arXiv:2507.03724 (v4), 2025.

[5] D. Kerestecioglu, A. Robsky, C. Vasters, A. Sharma, and Y. Kesselman. Human-Inspired Memory Architecture for LLM Agents. arXiv preprint arXiv:2605.08538, 2026.

Questions for authors

Q1. How large is the reconsolidation gain when no gold information is used? Please report both benchmarks with the reference answer and correctness label removed from the feedback. For LongMemEval, please also use warm-ups drawn from sessions that don't contain the evidence and replace the Prompt B.1 examples with questions outside the test set. Give the mean ± std over at least three runs, including runs without reconsolidation.

Q2. Which question order is reported in Tables 1 and 3, and why? The reported values match the Seed=2026 order in Fig. 3, the most favourable of the four. Averaged over the four orders, the gain is about +1.1 rather than +2.01.

Q3. For each baseline in Tables 1–2, what are the source, judge model, judge prompt and number of questions? Where does the Zep row come from? Can REALM, MemOS [4] and the other omitted systems be compared under a single judge and question set?

Q4. What exactly does the answer model see? Do edges added during reconsolidation get text descriptions, and are these written with the gold answer in view? Fig. 6 credits most of the gain to better use of the evidence: 15 more correct answers, but only 4 more questions where the evidence was found. Edge descriptions also enter retrieval scoring (Eq. 8).

Q5. Several details needed to judge cost and reproducibility are missing. Could the authors: a) report the number of LLM calls, tokens and latency per ingested turn and per question, compared with the baselines, since the paper claims REALM is "efficient"; b) state which benchmark Table 4 and Fig. 6 use; c) name the embedding model and decoding settings; d) provide the answer-generation prompt?

4 comments

Thank you for looking closely at our experimental setup and comparisons. One of your central questions is whether REALM can still organize its memory effectively when the updater has no reference answer. On LoCoMo, overall accuracy held up both with correctness labels alone and with neither form of feedback. We also re-scored the LoCoMo answers with MAGMA’s published rubric and the LongMemEval answers with its released evaluation code. We discuss those results and address your questions about the evaluation set, answer prompts, and cost.

[W1/Q1 — What happens when we remove answer feedback?]

When REALM first organizes its memory, it has yet to see what users will ask. A query can bring together pieces of information that were stored separately but are useful for the same task. REALM looks at the retrieved content and its topic to decide which local connections to record for later queries. As you pointed out, we supplied reference answers to the updater in the original experiments, and these may be unavailable in practice. We therefore tested how much the method depends on this feedback.

We ran two ablations on the same 1,540 answerable LoCoMo questions. One keeps the correctness label and removes the reference answer. The other removes both. The table reports the mean and sample standard deviation over four question orders, including the historical original order and three shuffles with seeds 0, 42, and 2026. It includes binary accuracy and MAGMA’s graded score on a 0–100 scale.

Updater inputsBinary accuracy (%)MAGMA score
GT + correctness74.84 ± 0.8871.19 ± 0.51
Correctness only76.10 ± 0.3771.74 ± 0.27
Neither75.55 ± 0.6371.57 ± 0.24

The mean with full feedback is 74.84%. The historical runs in W1/Q2 averaged 74.98%. The two means differ by about 0.15 percentage points, less than the variation across question orders in either set of runs. Both ablations yield slightly higher mean LoCoMo accuracy than the full-feedback setting. In the no-feedback setting, the judge still uses the reference answer for scoring. The updater receives no reference answers, correctness labels, or annotated evidence IDs.

You also ask how much reconsolidation helps without feedback. Relative to the paper’s single archived run without reconsolidation (73.96%), the four-order means are higher by 0.88 percentage points with full feedback, 2.14 with correctness labels alone, and 1.59 with no feedback. The sample standard deviations describe how the scores vary across the four question orders. The largest accuracy SD is 0.88 percentage points, in the full-feedback setting.

[W1 / Q2 — Question order and gain size]

Tables 1 and 3 report seed 2026, the best-performing order, with a gain of 2.01 percentage points. In the archived runs with full feedback, accuracy was 74.48% for seed 0, 74.48% for seed 42, 75.00% for the original order, and 75.97% for seed 2026. The mean is 74.98%, about 1.02 percentage points above the paper’s 73.96% reference without reconsolidation, close to your estimate of about 1.1 points. We will report all four results in the revision so readers can see how the gain varies with question order.

[W1 — Why we added warm-up questions to LongMemEval?]

Each LongMemEval sample pairs a separate history with one target question. To see whether earlier queries could improve the memory used to answer that question, we generated warm-up questions from selected sessions and let the system retrieve information and update its memory before answering the target. In this setup, the number of correct answers rose from 296 to 306 out of 470. You identified two details that matter for interpreting this result. We used annotations to select sessions containing evidence, and the prompt for generating warm-up questions included examples from the test set. The generator did not receive the current target question. We will make both details explicit in the revision and report the gain of 10 correct answers under this setup.

[W2/Q3 — Scoring, evaluation sets, and baseline sources]

Scoring. You’re right that Section 3.1 should distinguish the scoring methods. REALM’s original LoCoMo judge gives a binary verdict, while MAGMA gives graded credit for semantic matches. We will correct Section 3.1 and withdraw the 7.17-point comparison across rubrics. We re-scored the saved REALM answers using MAGMA’s published rubric with GPT-4o-mini at temperature 0. The four-order means are 71.19 with full feedback, 71.74 with correctness labels alone, and 71.57 with no feedback. Weighting the category scores in MAGMA’s Table 1 by the numbers of answerable questions gives about 68.8 for MAGMA and 68.7 for its Nemori entry. Under the same rubric, REALM scores about 2.4–3.0 points higher than these published references. Retrieval budgets and API versions may still differ.

Evaluation sets. As Appendix B.1 explains, we evaluate 1,540 answerable questions on LoCoMo and 470 on LongMemEval. We focus on whether the system can find existing evidence and answer from it. The unanswerable questions assess when to abstain. We also re-scored the same saved answers with MAGMA’s released LongMemEval code (evaluate_lenient, threshold 0.5). The system answered 291/470 (61.9%) correctly with warm-ups and 289/470 (61.5%) without them. The original binary judge gave 296/470 (63.0%) to the version without warm-ups. The warm-up setup described in W1 also applies to the re-scored result with warm-ups.

Baseline sources. The table shows which results come from our runs and which come from published comparisons. We will mark any missing details about the published settings so readers can judge which numbers are directly comparable.

LoCoMo rowSourceScoring / provenance
REALMOur saved answers, 1,540 answerable questionsBinary judge; separately re-scored with MAGMA's released rubric (GPT-4o-mini, temperature 0)
MIRIX, Mem0, ZepMemOS Table 3Literature results from its GPT-4o-mini comparison; not independently re-judged by us
MAGMA, NemoriMAGMA Table 1Literature graded scores; answerable-only averages recomputed from four rounded category scores
A-MemOur reported reproductionOriginal comparison under our evaluation setup; not a fresh multi-run reproduction

As you note, REALM leads in two of the six LongMemEval categories, and the results vary across categories. Our main metric is overall accuracy on the answerable questions. We report both overall and category-level results so readers can see where the gains occur.

We will add MemOS’s binary accuracy of 75.80% on LoCoMo and the LongMemEval results from its Table 4, which reports 66.4 for Mem0, 72.4 for Memobase, and 77.8 for MemOS. The LoCoMo comparison in W2/Q3 uses the same scoring rubric. For LongMemEval, we will present the published results with their original question sets and scoring methods.

We will add these works: HiMem, Human-Inspired Memory Architecture (Kerestecioglu et al., 2026, your reference [5]), Memory-R1, and ReasoningBank, and clarify what they update. It is important to note that Memory-R1 uses PPO or GRPO on training QA pairs to learn memory operations. REALM uses prompts to choose local relation edits without training the updater’s parameters. ReasoningBank distills reasoning strategies that can be used in later prompts. REALM edits connections between episodic memories, changing which paths later queries can follow to find evidence.

[Q4 — What the model sees and how to read Figure 6]

As you point out, the updater can revise edge descriptions, and those descriptions can affect later retrieval through the similarity term in Eq. 8. The answer model sees the content and descriptions of retrieved nodes, including original session text on LongMemEval. Edge descriptions affect what the model sees indirectly through retrieval. The updater itself receives the question, the generated answer, retrieved and seed nodes, existing edges, and any feedback allowed in that setting. This indirect path is why we tested in W1/Q1 whether accuracy holds up after removing feedback. The gold_supporting_node_ids field records nodes identified by the model itself. Annotated evidence IDs are never passed to the updater through this field.

In Figure 6, expansion retrieved annotated evidence for four more questions, and the number of correct answers in that group rose by 15. Even when the model finds the same direct evidence, it may answer better if it also sees useful supporting information or fewer distractors. The figure is consistent with that explanation. In the revision, we will specify how the figure measures evidence retrieval and answer correctness.

[Q5 — Cost, benchmarks, model settings, and answer prompts]

(a) Cost. We counted the model calls in the implementation. Graph traversal, scoring, and reranking run locally, as does embedding with all-MiniLM-L6-v2. For each turn, one call extracts U memory units. Each unit then takes one call to decide what to do with it and an additional call when there are candidate links, for a total of 1 + U to 1 + 2U calls. At question time, one call plans seed retrieval, and each expansion round can use up to two calls, first to check whether the evidence is sufficient and then to plan expansion for all selected seeds together. There is one call to generate the answer and one more when reconsolidation is enabled.

StageLLM callsNotes
Ingestion, per turn1 + U to 1 + 2U1 call generates U memory units; then 1 operation decision per unit, plus 1 edge-connection decision per unit only when link candidates exist
Answering, per question4–91 seed-planning call + 2 calls per expansion round (sufficiency check + expansion planner, all frontier seeds batched into one call) with early stopping, R ≤ 3 rounds; +1 answer generation; +1 reconsolidation update (omitted in the no-reconsolidation control)

With a budget of 64 recalled nodes and at most three expansion rounds, the system makes at most nine model calls per question when reconsolidation is enabled. If the initial seeds are enough at the first sufficiency check, it makes four. We did not retain complete token counts or latency measurements for the original runs. We will remove the efficiency comparisons that lack measurements and include directly measured token counts and latency in the revision.

(b) Benchmarks. Table 4 and Figure 6 both use LoCoMo. Table 4 covers 369 questions that require graph expansion. Path search, beam search, and adaptive search answer 147, 159, and 160 correctly, respectively. In Figure 6, “Found” means the supporting evidence was retrieved during expansion, and “Only” means it was retrieved exclusively during expansion. We will make the dataset and raw counts stated in the text explicit in the table and figure.

(c) Model settings. We use GPT-4o-mini at temperature 0 to generate answers and propose topology updates. We use all-MiniLM-L6-v2 locally for embeddings.

(d) Answer prompts. We include both templates used in the experiments. The answer context contains the retrieved node text and descriptions. The template covers categories and the two instructions after it are appended to temporal questions. The system message is “You are a conversational agent answering questions based on retrieved memory.”

# CONTEXT:
You have access to facts and entities from a conversation.
# INSTRUCTIONS:
1.Carefully analyze all provided memories
2.Pay special attention to the timestamps to determine the answer
3.If the question asks about a specific event or fact, look for direct evidence in the memories
4.If the memories contain contradictory information, prioritize the most recent memory
5.Always convert relative time references to specific dates, months, or years.
6.Be as specific as possible when talking about people, places, and events
7.Timestamps in memories represent the actual time the event occurred, not the time the event was mentioned in a message.

Clarification:
When interpreting memories, use the timestamp to determine when the described event happened, not when someone talked about the event.

Example:

Memory: (2023-03-15T16:33:00Z) I went to the vet yesterday.
Question: What day did I go to the vet?
Correct Answer: March 15, 2023
Explanation:
Even though the phrase says "yesterday," the timestamp shows the event was recorded as happening on March 15th. Therefore, the actual vet visit happened on that date, regardless of the word "yesterday" in the text.

# APPROACH (Think step by step):
1. First, examine all memories that contain information related to the question
2. Examine the timestamps and content of these memories carefully
3. Look for explicit mentions of dates, times, locations, or events that answer the question
4. If the answer requires calculation (e.g., converting relative time references), show your work
5. Formulate a precise, concise answer based solely on the evidence in the memories
6. Double-check that your answer directly addresses the question asked
7. Ensure your final answer is specific and avoids vague time references

Context:

{context}

Question: {question}
Answer:
- For "when" questions, you MUST answer with an exact time format (e.g.June 5 2020) or a relative time based on the exact time (e.g., two days after June 5).
- For other time-related questions, answer directly.

The timestamp in LongMemEval is the session time. If the user says on March 15 that they went to the vet “yesterday”, the answer is March 14. We have yet to re-evaluation with a corrected template. We will include the above in the revised version.

Thank you once again for taking the time to review our paper and for your thoughtful feedback. Your questions prompted additional feedback ablations across four question orders, re-scoring with MAGMA’s rubric, and clarifications of the evaluation setup. We sincerely hope that these additions address your main concerns. If you have any further questions or suggestions, we would be glad to offer further clarifications. Thank you for your time and consideration!

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

This paper proposes REALM, a long-term memory framework for LLM agents inspired by memory reconsolidation in cognitive neuroscience. The central idea is to allow previously stored memories to evolve through retrieval and subsequent task feedback, rather than updating memory only when new information arrives. REALM consists of three components: an autonomous memory organization module that constructs a heterogeneous cognitive graph containing different types of memory nodes and relations, an adaptive retrieval module that dynamically combines graph-search strategies, and a retrieval-driven reconsolidation module that modifies the activated memory subgraph by adding, strengthening, or weakening relations based on retrieved evidence and task feedback. The authors evaluate REALM on LoCoMo and LongMemEval using GPT-4o-mini, reporting overall accuracies of 75.97% and 65.11%, respectively. Ablation studies indicate that reconsolidation improves performance by approximately two percentage points on both benchmarks.

Critical strengths
  • Its studied problem in long-term agent memory by exploring how memory retrieval and subsequent task interactions can influence future memory organization is interesting.
  • The proposed framework provides a reasonably coherent system design for studying usage-driven memory evolution.
  • The experiments offer some empirical evidence that modifying memory relations after retrieval can improve downstream question-answering performance.
Critical weaknesses
  • The methodological novelty appears limited relative to existing agentic memory and graph-based retrieval systems. In particular, existing methods such as MAGMA [1] already incorporate structured graph memory and adaptive retrieval, while HeLa-Mem [2] explicitly leverages memory usage signals to improve future memory access. The paper does not convincingly establish why this particular formulation constitutes a substantial methodological advancement over existing usage-driven memory adaptation approaches.
  • The experimental protocol introduces a significant concern regarding additional supervision and evaluation validity. As described in Section 2.4 and Appendix B.4, the reconsolidation module receives not only the generated answer and correctness signal but also the ground-truth answer after each evaluated question. This provides the memory system with privileged information that would not generally be available during ordinary long-term agent interactions.
  • The experimental evidence provides relatively weak support for the central claim that reconsolidation substantially improves long-term memory organization. According to Table 3, reconsolidation improves overall performance by only 2.01 and 2.13 percentage points on LoCoMo and LongMemEval, respectively, whereas the majority of REALM's performance is already achieved without reconsolidation.
  • The analysis in Figure 6 attributes improvements to more informative memory organization, but the reported measurements primarily concern evidence retrieval and answer correctness rather than directly quantifying structural quality or demonstrating that the learned relations capture more accurate or reusable knowledge.

[1] Jiang et al. MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents. ACL 2026

[2] Zhu et al. HeLa-Mem: Hebbian Learning and Associative Memory for LLM Agents. ACL 2026

Questions for authors
  • What information in the ground-truth answer is actually used to determine the reconsolidation operations, and whether the method remains effective when only a binary correctness signal or imperfect user feedback is available?
  • How does the framework distinguish semantic relationships that should be permanently encoded in memory from temporary correlations arising from co-retrieval for a particular query?
  • To what extent does reconsolidation provide transferable improvements beyond the topics and questions that originally triggered the updates? Could the authors evaluate whether topology updates induced by one subset of questions improve performance on held-out questions with different reasoning requirements, or whether the gains diminish when subsequent queries are semantically unrelated to previous interactions?
  • What is the relative importance of the three reconsolidation operations, and how does the system determine whether an incorrect answer resulted from missing evidence, incorrect reasoning, or misleading memory relationships?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.