acceptodds
Under review as a conference paper at ICLR 2027

SLICE-Mem: Salience-Gated Lightweight Information Curation for Efficient LLM Agent Memory

Abstract

Long-term memory enables large language model (LLM) agents to maintain personalized and temporally coherent interactions, but repeated LLM inference for memory admission, extraction, and organization introduces substantial cost and latency. We present SLICE-Mem, a lightweight memory framework that combines salience-guided admission and atomic information extraction. A compact classifier identifies user turns worth retaining, and a semantic-dependency-based extractor converts admitted turns into concise memory entries. This online pipeline requires no generative LLM calls, while cross-entry reconciliation is deferred to an offline consolidation stage. We evaluate SLICE-Mem on the weekly subset of Memora and LongMemEval-S using 8B and 30B Qwen backbones for memory construction and consolidation on both benchmarks. Compared with the generative-memory baseline MemOS under the Qwen3-30B configuration, SLICE-Mem’s online pipeline achieves 46.60 versus 46.68 FAMA (Forgetting-Aware Memory Accuracy) on Memora Weekly and 56.8% versus 59.3% accuracy on LongMemEval-S, while reducing total LLM token consumption by 92.4% and 99.5%, respectively. These results demonstrate that SLICE-Mem can preserve useful evidence for long-term memory with substantially less reliance on LLM-based memory construction.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agent Memory, Salience-Guided Memory Curation, Efficient Memory Construction

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_K1dpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Previous:

SLICE-Mem gates memory writes with a salience score from LLM teachers, keeping only high-salience entries. It reports 92.4%–99.5% token reductions with largely unchanged task metrics. Main concerns: the gating threshold τg=0.2 was selected on the test benchmarks; the Mem0 baseline (13.8%) is far below public numbers; and the offline consolidation variant yields negative marginal gains.

Update:

This paper, SLICE-Mem, propose a compact salience classifier with the extraction operation to construct memories without online generative LLM calls. Optional offline consolidation handles cross-entry updates. Besides, the paper studies the trade-off between memory quality and inference cost on several benchmarks.

Critical strengths

Previous:

Large, direct efficiency gains:1–2 orders of magnitude token reduction with flat task metrics — clear engineering value. Simple and reproducible:Intuitive mechanism with a fully disclosed teacher-annotation pipeline (Appendix A.5).

Update:

Moreover, avoiding generative calls during online construction offers clear practical value. The revised evaluation and analysis prove that the trade-off between quality–cost.

Critical weaknesses

Previous:

Threshold tuned on test data:τg=0.2 is selected on the evaluation benchmarks without an independent validation split, risking evaluation overfitting. Abnormal baseline and unclear attribution:1Mem0 at 13.8% diverges from public results; the offline variant's negative gain obscures which component is effective.

Update:

The threshold sweep shows a relatively stable range, but the validation for selecting the threshold is not stable and solid. Besides, the extractor’s value is also limited. Teacher annotation adds about 8M tokens ( one-time cost), which is not included.

Questions for authors
  1. How should τg be chosen without touching the test set? Please provide a sensitivity curve over τg.
  2. Please explain the Mem0 13.8% baseline (version, configuration, protocol).
  3. Is teacher-annotation cost counted in the total budget, and how does it compare to the saved tokens?
  4. What are typical failure cases where the gate drops critical memories, and how do they affect long-horizon tasks?
2 comments

Thank you for drawing our attention to the unusually low Mem0 score, which we have corrected in Table 2 and Appendix D.2 (Table 7) after rerunning both model sizes. Appendix C.3 (Table 5) reports the full threshold sweep, and Section 5 discusses how missed memories can affect later answers.

R1. How can the threshold be chosen before testing, and how sensitive is performance to it? (W1, Q1)

  • We conducted a sensitivity analysis on Memora Weekly by varying the Gate threshold τg from 0.1 to 0.9 in steps of 0.1. As shown in Table 5 of Appendix C.3, FAMA was relatively stable across thresholds from 0.1 to 0.6 (46.81–48.54) and declined at higher thresholds, reaching 42.10 at 0.9. We used τg = 0.2 as a reference value within this range. We applied the same reference value directly to LongMemEval-S without further tuning.

Table 5: Sensitivity of memory curation performance to Salience Gate threshold

Gate threshold (τg\tau_g)0.10.20.30.40.50.60.70.80.9
Memora Weekly FAMA46.8148.5448.0747.3046.9946.8944.8744.1342.10

R2. Why did Mem0 score 13.8%, and what does offline consolidation contribute to SLICE-Mem? (W1, Q2)

  • In our original 30B run, the evaluation pipeline marked some failed memory writes as successful, leaving gaps in the memory stores used to answer questions. We corrected the pipeline and reran all LongMemEval-S questions with open-source Mem0 SDK v2.0.13, with Qwen3-8B and Qwen3-30B-A3B-Instruct-2507. The corrected accuracy was 54.4% (272/500) with 8B and 55.6% (278/500) with 30B. We have updated both in Table 2 in main text and Table 7 in Appendix D.2.
  • With the 30B consolidator, offline consolidation helped most when the system had to recall facts the user had shared earlier (the Remembering task on Memora Weekly). FAMA rose from 46.24 to 50.55 (Table 6) as filtering out duplicate or outdated entries can make current facts easier to find. However, an older entry may still contain a preference that matters for a future recommendation. Making that entry harder to retrieve may help explain why the same system scored lower when asked to recommend something suited to the user (the Recommending task). We discuss this trade-off in Section 5.

R3. Is teacher annotation included in the token budget? (Q3)

  • The main tables report operating costs and exclude the one-time cost of teacher annotation. Annotation for the training examples alone accounted for 8.585M tokens. Once trained, the Gate can be reused across conversations, and online memory admission and extraction require no generative LLM calls. On Memora Weekly, SLICE-Mem used 6.473M fewer tokens than the 30B MemOS baseline. These savings accumulate with each workload, while the annotation cost is incurred only once.

R4. What are typical failure cases and how missed turns matter for long-term use? (Q4)

  • Context-dependent turns and included vague or ambiguous turns are the typical failure cases that Gate treat as false positive/negative.
  • If that turn is the only record, retrieval cannot recover the omitted detail. A later answer may overlook a preference or rely on an outdated task status. Generative memory systems can also lose details during summarization. The end-to-end scores in Tables 1–2 capture a loss when it changes a graded answer. We discuss these limits in Section 5 (discussion and limitation).

Thanks authors for correcting the Mem0 evaluation and adding the threshold and cost analyses. I have updated and withdraw the broad criticism based on the original 13.8% score.

Besides, the sweep partialy solves my concern about threshold sensitivity, but the threshold selection os still unclear. The consolidation results also clarify a task-dependent trade-off rather than a uniformly negative effect.

Overall, My remaining concerns are the threshold-selection protocol and the extractor’s limited benefit. I will keep my score of 2.

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

This paper resolves the repetitive call of llm inference when generating or updating the long term memory. The author proposes a light weight memory frame work called SLICE-MEM composed by the core components: compact classier which identifies user turns worth retaining, and semantic dependency-based extractor which converts admitted turns into memory entries. SLICE-Mem is evaluated on Memora Weekly and LongMemEval-S using 8B and 30B Qwen backbones for memory construction and consolidation. Experimental results show competitive downstream accuracy compared with existing memory frameworks, while reducing total LLM token consumption by 92.4% and 99.5%, respectively, compared with MemOS under the 30B configuration.

Critical strengths
  • The observation that the long term memory requires a lot of llm inference which costs a lot is realistic overhead
  • For offline memory generation track, the result is compelling
  • Detailed ablation, scan of the parameter etc. and I appreciate author is honest about the limitation of consolidation
Critical weaknesses
  • Deterministic extractor mechanism's necessity comparing to naive generation is not proved
  • The salince gate have complete information while the extractor may not have the information to restore the complete context e.g. A.5 example. According to author's rule it worths being memory and depend on user message but the extractor can only see the User part which can't claim that the correct extraction means self contained memory is produced.
Questions for authors
  • Can the authors compare SLICE-Mem against a Gate + Raw Turn baseline to clarify whether dependency-guided extraction provides a actual benefit beyond salience-based filtering, under comparable storage and retrieval budgets.
  • How does the system preserve the meaning of context-dependent user turns when the extractor only receives the target turn? A focused analysis of omitted referents, incomplete propositions, and their downstream impact would help establish the scope of the evidence-preservation claim.
  • Can the authors clarify the evaluation metric and variability? Why does the 30B MemOS result appear consistent with 499 rather than 500 evaluated questions?and how stable are the main effectiveness comparisons across evaluation runs?
1 comment

Thank you for proposing the Gate + Raw Turn comparison, which we added to Appendix D.4 (Table 9). Section 5 now explains what information can be lost when the extractor sees only the user’s reply. We also completed the missing 30B MemOS question and corrected its accuracy in Table 2.

R1. What does deterministic extraction offer over generative extraction? (W1)

  • The w/o Extractor ablation in Table 1 keeps the Gate and uses an LLM to extract memories from admitted turns. On Memora Weekly, deterministic extraction scored 46.60 FAMA, compared with 48.02 for 30B generative extraction and 44.05 for 8B.
  • Deterministic extraction used zero generative calls during online construction. The 30B extractor used 1351 calls and 0.746M tokens, and the 8B extractor used 1345 calls and 0.729M tokens. Table 1 reports construction times of 0.362, 0.442, and 0.358 h for deterministic, 30B, and 8B extraction, respectively.
  • We chose deterministic extraction for SLICE-Mem because it achieved FAMA scores comparable to those of LLM-based extraction while eliminating all generative LLM calls and tokens from memory construction.

R2. How does extraction compare with storing the admitted turn verbatim? (Q1)

  • Both systems (SLICE-Mem vs Gate + Raw Turn) used the same frozen Gate, user turns, embedding model, retrieval budget, answer model, and judge. Table 9 shows that Gate + Raw Turn scored higher on both benchmarks with similar QA token use compared to SLICE-Mem.
  • We retain the extractor as an option that want memories written as explicit statements about the user. Across the stores used for QA, extraction reduced the total amount of indexed text by 5.5% on Memora Weekly and 5.3% on LongMemEval-S. This gives a concrete reason to use the extractor when the form and size of the stored memories matter. We have updated the information in Appendix D.4.

Table 9. Evaluation on QA performance, indexed entries, and token use with and without extraction on Memora Weekly and LongMemEval-S.

BenchmarkMethodQA scoreIndexed entriesQA tokens (M)
Memora WeeklyGate + Raw Turn51.12 FAMA33270.534
Memora WeeklySLICE-Mem46.97 FAMA32300.538
LongMemEval-SGate + Raw Turn56.6%989870.914
LongMemEval-SSLICE-Mem55.4%894110.913

R3. Can a memory preserve meaning when the extractor sees only the user’s reply? (W2, Q2)

  • The Gate reads the user’s message and the preceding message to decide whether to keep the message. The extractor then reads only the reply. We chose user messages as the source of memory because Memora focuses on what users expect the system to remember. This makes it easier to compare our work with the baseline.
  • We acknowledge that extracting from a single turn can leave a memory incomplete or ambiguous. The extractor may omit a detail that the user stated explicitly, or it may retain a reference that only makes sense in light of an earlier message. In either case, the stored memory can lack information needed to interpret the user’s meaning, which can lead to an incorrect answer later. We clarify this limitation in Section 5 (Discussion and Limitations).

R4. How are the scores computed, and why did 30B MemOS have 499 questions? (Q3)

  • A construction service failure caused us to skip one question in the original 30B MemOS run, giving 296/499 (59.3%). We rebuilt that history from an empty store and reran retrieval, QA, and grading under the original settings. MemOS answered the question incorrectly. The corrected Table 2 result is 296/500 (59.2%).
Reviewer_kxXCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a technique to use a trained filter to decide admission of the memory based on memory content. Then it is processed with deterministic parser to extract syntactical information from the content and store into the memory. The memory system is overall similar to existing graph-based memory systems (updates, retrieval etc). Overall, the trained gate is a simple and effective way to filter trivial content that might add noise to the system, and the evaluation shows its effectiveness.

Critical strengths
  1. The main idea is to train a gate filter from LLMs so that it can be used to filter trivial memory items before going into the memory system. The training approach is effectively distilling from LLM generated responses, and it is cheap to apply in practice.
Critical weaknesses
  1. Other parts of the paper seems relatively standard memory system. My main concern here is simply that the gate mechanism is general, and I'm not sure why it is only applied in the specific designed memory pipeline rather than evaluating on top of existing memory system? For example, I'm curious what if the trained gate is directly applied to previous system to see if the save in memory processing time can help them improve overall effectiveness while reducing cost?

  2. Some qualitative analysis on what type of memory items are likely getting filtered as false positive/negative would help understanding whether a simple gate like this is generalizable.

Questions for authors

Whether the gating is effective for more generic settings.

2 comments

Thank you for suggesting the transfer experiment that we now report in Appendix D.3 (Table 8). We tested the same Gate in five existing memory frameworks and added analysis of the turns it mistakenly retains or discards in Appendix C.3.

R1. Does the Gate benefit other memory frameworks? (W1, Q1)

  • We applied the same frozen Gate to five frameworks on Memora Weekly with the 30B construction backbone. Within each pair, the Gate filtered user turns. We preserved the framework’s handling of assistant messages and used the same retrieval, answer model, and judge.
  • All five frameworks used fewer LLM tokens. A-MEM and LangMem Manager gained 5.09 and 6.19 FAMA points, respectively. FAMA fell for the other three frameworks. Only MemOS and LightMem took less time to construct and update their memories.
  • These results suggest that whether filtering improves answer quality or saves time depends on how each framework constructs and memory mechanism. The result is shown in the Table 8 and we have updated in the Appendix D.3.

Table 8. Effect of Salience Gate on answer quality and resource use across five memory frameworks on Memora Weekly. (Each cell reports without Gate → with Gate. The symbol ↑ indicates an increase in value, and ↓ indicates a decrease in value. Total LLM tokens and LMM call counts cover memory construction, offline updates, and answer generation. Memory-processing time is measured in hours and covers construction and offline updates, including Gate inference when the Gate is used.)

FrameworkFAMATotal tokens (M)LLM callsTime (h)
A-MEM35.01 → 40.10 ↑11.814 → 9.834 ↓3269 → 3272 ↑3.647 → 4.028 ↑
LangMem Manager26.59 → 32.78 ↑18.763 → 17.222 ↓3081 → 3125 ↑6.633 → 7.891 ↑
Mem048.74 → 47.24 ↓15.487 → 15.219 ↓1785 → 1781 ↓1.830 → 1.929 ↑
MemOS47.01 → 46.60 ↓7.631 → 6.630 ↓5801 → 5239 ↓2.826 → 2.701 ↓
LightMem38.16 → 33.95 ↓20.145 → 17.038 ↓24344 → 19969 ↓9.023 → 6.662 ↓

R2. Which turns does the Gate mistakenly retain or discard? (W2)

  • At the deployed threshold of 0.2, the Gate retained 1608 of 1613 Weekly user turns labeled for retention (99.69% recall). It also retained 1719 turns labeled negative (48.33% precision).
  • False negatives tended to be context-dependent turns: they added new information, but the reader needed an earlier turn to identify what the user was referring to or what had changed. False positives included vague or ambiguous turns: the Gate retained them even though they added little information beyond what the conversation had already established. These results support the Gate’s ability to generalize across the dataset. The information was updated in the Appendix C.3.
Reviewer_kxXCOpenReview ↗

Interesting result!

Really appreciate the reports on transferability of the gate across systems. I think the trade off between quality and token efficiency is quite interesting there, this might also mean existing memory systems have different weaknesses.

I think the paper would be much stronger in future if a transferrable (and lightweight) filtering machanism can be developed. The current paper has some limitations (I'm keeping score 2), but I think it would be very promising in future.

Comments on the paper

1 comment

We thank the Area Chair and all three reviewers for their careful reading of SLICE-Mem. Your questions led us to test which parts of the framework account for its savings, when filtering changes answer quality, and how reliably the baselines were evaluated. We added two comparisons in Appendices D.3 and D.4, corrected the LongMemEval-S results in Table 2, and expanded the discussion in Section 5.

  1. Does the Gate benefit other memory frameworks? (kxXC W1, Q1) We applied the same frozen Gate to five frameworks on Memora Weekly with the 30B construction backbone. In each pair, the Gate filtered user turns, with the same handling of assistant messages, retrieval settings, answer model, and judge. All five used fewer LLM tokens. A-MEM and LangMem Manager gained 5.09 and 6.19 FAMA points, respectively, while the other three scored lower. Only MemOS and LightMem took less time to construct and update their memories when Gate inference was included (Appendix D.3, Table 8).

  2. What does deterministic extraction add? (yihL W1, Q1) With the Gate fixed, deterministic extraction scored between the 8B and 30B generative extractors on Weekly and required no generative calls during online construction (Table 1). Storing admitted turns verbatim (Gate + Raw Turn) scored higher on both benchmarks with similar QA token use under the same embedding, retrieval, answer, and grading settings. The extractor remains an option for writing explicit statements about the user. It reduced the amount of indexed text by 5.5% on Weekly and 5.3% on LongMemEval-S (Appendix D.4, Table 9).

  3. What information can be lost? (kxXC W2, yihL W2 and Q2, K1dp Q4) At the deployed threshold, the Gate retained 1608 of 1613 Weekly turns labeled for retention, with 99.69% recall and 48.33% precision. Missed turns tended to add information whose meaning depended on an earlier message. Mistakenly retained turns included vague replies that added little new information (Appendix C.3). The extractor reads only the user's reply, so it may preserve a reference without the context needed to understand it, or omit a detail the user stated explicitly. If this was the only record, a later answer may overlook a preference or rely on an outdated task status (Section 5).

  4. How sensitive are the results to the Gate threshold? (K1dp W1, Q1) We evaluated nine thresholds from 0.1 to 0.9 on Memora Weekly. FAMA ranged from 46.81 to 48.54 for thresholds between 0.1 and 0.6, then declined at higher thresholds, reaching 42.10 at 0.9. We used 0.2 as the reference setting and applied it directly to LongMemEval-S without further tuning. Appendix C.3, Table 5 reports the full sweep.

  5. Why did the LongMemEval-S baseline results change? (yihL Q3, K1dp W1 and Q2) In the original 30B Mem0 run, the evaluation pipeline marked some failed writes as successful, leaving gaps in the stored histories. We corrected the pipeline and reran all 500 questions with Mem0 SDK v2.0.13. Accuracy was 54.4% with 8B and 55.6% with 30B (Table 2 and Appendix D.2, Table 7). For 30B MemOS, a construction service failure had left one question unevaluated. We rebuilt its history from an empty store and reran retrieval, QA, and grading under the original settings. The answer was incorrect, bringing the result to 296/500 (59.2%) in Table 2.

  6. When does offline consolidation help? (K1dp W1) On Memora Weekly, offline consolidation in the 30B setting raised FAMA from 46.24 to 50.55 on questions that asked the system to recall facts the user had shared, while the score for personalized recommendations fell (Table 6). Filtering out duplicate or outdated entries can make current facts easier to retrieve. However, consolidation marks whole entries as outdated, and an entry may contain both a superseded fact and a preference that still applies. If that entry is harder to retrieve, the system may overlook the preference when making a recommendation. We discuss this trade-off in Section 5.

  7. How does teacher annotation affect the token budget? (K1dp Q3) The main tables report operating costs. Teacher annotation for the training examples accounted for 8.585M tokens as a one-time cost. On one Weekly evaluation, SLICE-Mem used 6.473M fewer tokens than 30B MemOS. The trained Gate can be reused across conversations, and online admission and extraction require no generative LLM calls, so the operational savings accumulate with use.

We thank the reviewers again for suggesting these comparisons and the Area Chair for considering the revised results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.