acceptodds
Under review as a conference paper at ICLR 2027

Learning from Others, Acting for You: Cross-User Memory Sharing for LLM Agents

Abstract

Large language model (LLM) agents serving different users often solve related tasks, yet separate user histories can leave reusable experience inaccessible to other agents. Pooling memories expands access but risks transferring preferences that conflict with the receiving user's requirements. We introduce ShareMem, a memory architecture that shares reusable experience while grounding its application in the receiving user's own preferences. Shared experiences indicate how to act and which preferences to consult; the receiving user's memory supplies their concrete values. Two-stage consolidation refines experience locally before integrating accepted edits into a shared pool. During execution, scope-first retrieval jointly selects local and shared experiences under a common entry budget, while a user-bound channel supports initial and agent-initiated preference retrieval. We evaluate ShareMem across web navigation (Mind2Web), online personalized interaction (VitaBench 2.0), and multi-session coding (MemoryCode) with four backbone models. It improves step success, average task success, and dialogue-macro coding scores, respectively, over matched user-local memory across all four models. Ablations favor two-stage consolidation for smaller shared pools, lower induction token usage, and better downstream performance, and support complementarity between experience guidance and active preference retrieval. Further analyses show that sharing helps most when relevant local experience is scarce, while source quality and cross-user preference interference limit useful transfer.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Agent Memory, Experience Sharing, Personal Preferences, Cross-User Transfer

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CLhfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an approach that improves task performance through experience sharing among agents in multi-user deployments. Reusable experience acquired by one user's agent remains inaccessible to other users, while pooling memories risks transferring personal preferences. The authors introduce ShareMem, which shares only how to act and which preferences to consult. Source-specific values in experiences are abstracted into semantic slots, while concrete preference values are supplied by each user's own preference store. Experiences are first refined in a local pool by an LLM-based manager through add, replace, and remove operations, and a shared manager then independently updates the shared pool using only accepted local edits as evidence. During execution, scope-first retrieval selects local and shared experiences from closer scopes first under a joint entry budget, alongside initial and agent-initiated preference retrieval. Across Mind2Web, VitaBench 2.0, and MemoryCode with four backbone models, the authors report that ShareMem outperforms user-local memory across all backbones.

Critical strengths
  1. The ShareMem design — separating shared memory and personal preferences is a reasonable structure for sharing the agent experience from other users while preserving the user's personal preferences.

  2. Ablation study gives further understandings — Two-stage consolidation, which updates the shared pool using only accepted local edits rather than raw sessions, yields a more compact shared pool and substantially lower induction token usage than direct writing, while also improving downstream performance.

Critical weaknesses
  1. The central premise of ShareMem remains unverified. The paper does not define what constitutes a user-specific value as opposed to reusable procedure, leaving this boundary to the LLM manager's judgment, and the separation relies solely on prompt instructions (Appendix C) without any audit of the shared pool for leaked values. Given the paper's own finding that agents misuse other users' preferences despite explicit instructions (Tables 16–17), it is unclear whether ShareMem actually prevents cross-user preference transfer.

  2. The evidence for the core claim is limited. Mind2Web involves no preferences and assigns each website's data to a single user, and MemoryCode's shared pools are minimal (40, 17, and 1 entries; Table 4), leaving VitaBench 2.0 as the main test, where the gains are modest. Moreover, no results report variance or significance tests (Tables 1–3), so the reliability of these small gains is unclear.

Questions for authors
  1. In the perspective of ShareMem's weakness mentioned in W2, especially for MemoryCode, shared pools contain only 40, 17, and 1 entries (Table 4), it is unclear whether sharing meaningfully contributes in this setting. How often are shared entries actually retrieved, and does ShareMem differ from LOCAL on subset 7, where only a single shared entry exists?

  2. The paper offers little qualitative evidence for ShareMem itself. What fraction of shared pool entries contain values traceable to contributing users, such as the examples mentioned in Table 17? Could the authors apply the audit protocol of Table 16 to ShareMem's failures, and provide examples of extracted shared experiences, cases where ShareMem succeeds but LOCAL fails, and the types of failures that remain?

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

The paper introduces ShareMem, an architecture for sharing procedural experience across users while resolving concrete preferences through the receiving user’s memory. It combines two-stage local-to-shared consolidation, scope-first retrieval under a joint experience budget, and initial and active preference retrieval. Experiments on Mind2Web, VitaBench 2.0, and MemoryCode with four backbones report improvements over matched user-local configurations on the principal metrics. Ablations and further analyses investigate consolidation, retrieval, experience coverage, source quality, and preference interference.

Critical strengths
  1. Using shared experience to guide which user preferences to retrieve is a useful aspect of the method. It connects reusable procedures with the information needed to personalize their execution.
  2. The work shows that sharing experience is most beneficial when relevant local experience is scarce, and that access to other users’ preferences can lead agents to apply inappropriate personal constraints.
  3. The paper is well organized and easy to follow.
Critical weaknesses
  1. The novelty over related work appears limited, as Collaborative Memory [1] already separates private and shared memory and removes user-specific details.
  2. Figure 4(b) adds local and shared experiences together, leaving the specific benefit of shared preference-retrieval guidance untested.
  3. Mind2Web has no preference channel, so its results support cross-user experience reuse but do not directly validate the proposed integration of shared guidance with user-specific preferences.
  4. The write-path ablations rebuild local pools and change extraction context, preventing clean attribution to shared-memory consolidation (§4.3).
  5. Closely related methods, including Collaborative Memory [1] and SkillClaw [2], are cited but not compared, without justification for their exclusion.
  6. The experiments lack statistical significance testing.
  7. The code is not released, which may lead to difficulties in reproducing the paper.

[1] Rezazadeh A, Li Z, Lou A, et al. Collaborative Memory: Multi-User Memory Sharing in LLM Agents with Dynamic Access Control. arXiv:2505.18279, 2025.

[2] Ma Z, Yang S, Ji Y, et al. SkillClaw: Let Skills Evolve Collectively with Agentic Evolver. arXiv:2604.08377, 2026.

Questions for authors
  1. What new mechanisms does ShareMem introduce beyond existing private/shared memory and de-personalization approaches?
  2. What evidence isolates the benefit of preference-retrieval guidance in shared experiences from that of local experience?
  3. Does shared consolidation improve over directly pooling the same frozen local experiences under matched retrieval and preference-access settings?
  4. How does ShareMem compare with Collaborative Memory or SkillClaw under matched conditions? If these methods cannot be suitably adapted, what prevents a fair comparison?
Reviewer_mwQ6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ShareMem addresses the isolation of reusable experience across personalized LLM agents. It combines two-stage memory consolidation, scope-first retrieval of local and shared experiences, and a user-bound preference channel with agent-initiated retrieval. Experiments on Mind2Web, etc datasets using diverse backbone models show improvements over user-local memory.

Critical strengths

The paper has broad empirical evaluations. Experiments span three benchmarks and four backbone models. Across these evaluations, ShareMem consistently improves over matched local-memory baselines, especially clear gains on Mind2Web. The ablations also provide useful insights into the roles of memory quality, coverage, and consolidation.

Besides, the paper clearly distinguishes reusable procedural knowledge from user-specific preferences, This is a clear motivation and address one core challenge in deploying personalized agents across multiple users.

Critical weaknesses

ShareMem introduces additional memory-update and retrieval operations, but the main evaluation does not provide its end-to-end latency, inference cost, or scalability relative to simpler alternatives.

Several ablations change multiple factors simultaneously. For example, the append-only comparison changes both memory maintenance and extraction context, while the active-retrieval ablation changes information access and interaction opportunities. Consequently, the individual contributions of the proposed components remain uncertain.

Code is not provided, raising concerns on the reproducibility.

Questions for authors

See the weakness section above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.