acceptodds
Under review as a conference paper at ICLR 2027

Demotion Beats Deletion: Graded Forgetting for Capacity-Bounded Agent Memory

Abstract

Persistent LLM agents accumulate memory faster than they can afford to read it, and every deployed system resolves this by deleting. Anthropic’s Memory Stores cap a store at 10,000 entries and instruct operators to remove stale ones; Mem0 prunes through LLM-issued DELETE; ChatGPT’s background consolidation rewrites in place. Deletion is an irreversible decision taken without knowing which memory a future query will need. We ask whether it is necessary. We present graded forgetting: traces are demoted through discrete retention stages and compressed to progressively lossier fidelity tiers, but never removed, so a capacity budget is met by shedding detail rather than items. We introduce recall-after-eviction, a protocol that isolates the capability at stake by conditioning on questions whose supporting evidence a capacity bound has destroyed. On an 833-day, 1,115-memory timeline held to 35% of its full size, deletion answers 2.5% of questions about the evicted past while graded forgetting answers 60.0% - 80% of the uncapped ceiling, from the same byte budget. Against Mem0, the strongest published baseline and one with no capacity policy of its own, the ordering inverts precisely when the budget binds: Mem0 leads without a bound and falls to 26.7% under one, against 66.7% for graded forgetting. The contribution is a capability under capacity pressure rather than an accuracy gain, and we argue this is the axis on which agent memory should be evaluated.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agent memory, LLM memory, Regenerative Memory

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_xz9hOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes graded forgetting for capacity-bounded LLM agent memory, where old memories are progressively compressed rather than deleted. It also introduces a recall-after-eviction evaluation protocol that specifically measures recall of information removed by a destructive retention policy. On a LoCoMo-based long-term memory benchmark under a 35% storage budget, graded forgetting substantially outperforms delete-oldest (60.0% vs. 2.5% accuracy) and also outperforms capacity-constrained Mem0 (66.7% vs. 26.7%).

Critical strengths
  1. The problem of long memory is real and challenging.
  2. The main empirical effect is large and easy to interpret.
Critical weaknesses
  1. The comparison against destructive forgetting is currently too weak to establish the broader case for graded forgetting. The decisive baseline is delete-oldest, which is a useful baseline, but it does not establish that graded forgetting is preferable to a realistic scenario. Cited papers are not properly discussed, such as MemRefine and "Learning What Not to Forget: Long-Horizon Agent Memory from a Few Kilobytes of Learning", which directly studies learned relevance-based eviction under matched budgets and includes LoCoMo evaluation. Comparisons against these methods, or at least comparable strong content-aware policies, seem necessary to determine whether the contribution is a competitive memory-management algorithm rather than a demonstration that compression can outperform recency deletion.

  2. The empirical validation is too narrow for the current claims, especially given the computational cost. The main result contains only 40 questions from one 833-day timeline, at one 35% storage budget, using one primary backbone. There is no budget sweep, no evaluation across multiple independent timelines or benchmarks, and no uncertainty estimate. This makes it difficult to know the shape or robustness of the claimed advantage.

  3. Token cost. Graded forgetting uses 33,423 tokens/query versus 2,806 for deletion, approximately a 12x difference. I agree with the authors that this may be an implementation artifact rather than intrinsic to compression, but as evaluated it is a substantial practical cost. Since the paper is motivated by resource-constrained deployment, storage cannot be considered entirely independently of inference cost.

Questions for authors

See Weaknesses.

Reviewer_dKKwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1(−1)
Soundness
1(−1)
Read review
Paper summary

This paper proposes "graded forgetting," a method in which agents compress old memories in retension stages progressively rather than deleting them entirely. Each memory is assigned a retention stage and a fidelity tier based on elapsed time; while details of older memories are lost, they are preserved as compact representations that can be reconstructed using contextual cues. The authors also introduce an evaluation protocol called "recall-after-eviction," which selects questions based on information that has been completely removed by the deletion policy. In an evaluation using 40 questions—conducted by concatenating three LoCoMo conversations (totaling 1,115 memories) and limiting storage capacity to 35% of the original size—graded forgetting achieved an accuracy rate of 60.0%, whereas an "oldest-first deletion" method achieved 2.5%. However, graded forgetting consumed 33,423 tokens per question, compared to 2,806 tokens for the deletion method.

Critical strengths

The paper presents the recall-after-eviction protocol to explicitly measures the ability to recover information under storage pressure, which is an interesting evaluation perspective.

Critical weaknesses

The evaluation is lacking:

  • The paper only compares to the deletion strategy and neglects the other baselines.
  • The paper only experiment in three conversations and one model with heuristic setup. The presentation is poor. eg the figures are not well prepared, the text is written in a fully AI style with seemingly no human care.
Questions for authors

How does the method perform compare to the other baselines, such as those listed in section 7? Can you show that the proposed method work in other proprietary and open source models?

Reviewer_cH21OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies memory retention under a storage budget. Instead of deleting old memories, it progressively compresses each trace and uses an age-dependent retrieval rule with a strong-cue override. Its recall-after-eviction protocol evaluates questions whose supporting evidence was completely removed by oldest-first deletion. On 40 such questions drawn from a concatenated LoCoMo timeline, graded forgetting achieves 60.0% accuracy versus 2.5% for deletion and 75.0% for an uncapped store. A separate 30-question experiment compares against Mem0 with an added oldest-first eviction rule.

Critical strengths

The paper asks a useful question that ordinary fixed-history benchmarks can miss: what remains answerable after storage actually becomes scarce? The analysis of duplicated evidence explains why random thinning may barely affect the questions being measured. The admission-rule example is also concrete. Showing that a product gate can make old memories unreachable, then adding a strong-cue override, makes the retrieval failure understandable. I appreciate the explicit reporting of the roughly twelvefold query-token overhead and the weaknesses of oldest-first deletion. These disclosures make the actual result easier to assess.

Critical weaknesses
  1. The capacity claim needs a precise boundary. Every retained memory still requires a nonempty scaffold, identifiers, and retrieval metadata. A fixed byte budget cannot hold an indefinitely growing number of individually addressable traces once all have reached their minimum tier. The method does not specify what happens at that point. Moreover, version history is described as preserving the original text 'for free'; its storage and any recovery access must be included or explicitly excluded from the budget. The 10,000-entry limit motivating the introduction is not solved by reducing bytes while retaining every entry.
  2. Recall-after-eviction is a useful diagnostic subset, but it conditions on a failure of the chosen deletion policy. It cannot by itself establish which policy allocates a fixed capacity better over the full query distribution. Compression may lose details needed by other questions. The paper needs overall accuracy, the fraction of evidence-lost questions, and conditional accuracy together, across several budgets and stronger retention policies.
  3. The comparison changes more than retention. Table 3 uses 33,423 tokens per query for graded forgetting versus 2,806 for deletion. The reconstruction loop may therefore contribute substantially to the gain. A shared retrieval/reconstruction procedure, plus a simple gist-only store at the same byte budget, would distinguish graded stages and cue gating from ordinary summarization and additional inference.
  4. The method is not specified well enough to reproduce from the manuscript. The tier contents, compression prompts, parameter values, contextual-vector construction, reconstruction procedure, and byte-accounting rules are missing. The 40-question main test and 30-question Mem0 test are too small and narrow to support the broad claims about deployed memory systems; Mem0's observed behavior should be attributed to this workload and configuration.
Questions for authors
  1. What exactly counts toward the byte cap: text, tags, graph edges, embeddings, metadata, and historical versions? Can any original text outside that cap be accessed during evaluation?
  2. What is the minimum footprint per retained trace, and what does the policy do when the sum of these minimum footprints exceeds the budget?
  3. Under a shared inference budget, how does the method compare with uniformly compressed gists, importance-based deletion, and a compression-plus-deletion policy on all 382 questions? Please also provide the tier-generation and recovery details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.