PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Jinming Hu, Haodong Zhao, Qi Jia, Die Chen, Tianhang Zhao, Sufeng Duan, Gongshen Liu

arXiv:2609.32511v1cs.AI

TL;DR

Separate user histories can isolate reusable experience, while pooled memories may expose preferences that conflict with the recipient’s requirements. ShareMem combines shared guidance with recipient-bound preference retrieval, and improves outcomes over matched user-local memory across three task domains.

  • Problem

    Separate histories can isolate useful experience, while pooling memories can introduce instructions that conflict with the receiving user’s requirements.

    Source: §1 · pp. 1–2

  • Method

    ShareMem consolidates reusable task experience into a shared pool, jointly retrieves local and shared guidance, and resolves concrete preferences through the receiving user’s memory.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    Across web navigation, online personalized interaction, and multi-session coding, sharing adds value beyond user-local memory; on Mind2Web, step success rises 2.35–4.10 percentage points over LOCAL.

    Source: §4.2 · p. 5, §5 · p. 10

  • Takeaways & Limitations

    The findings support distinguishing information access from applicability, since larger memory pools and broader access to other users’ preferences do not necessarily improve performance.

    Source: §5 · p. 10

    The evaluation fixes the receiving user and prevents cross-user preference access, within a setting of dedicated agents serving users with potentially similar tasks and differing requirements.

    Source: §3.1 · p. 3, §3.3 · p. 4

Abstract

from arXiv · show

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.

1 INTRODUCTION

ShareMem addresses the tension between making reusable experience accessible and preventing other users’ preferences from overriding the receiving user’s requirements. It shares task guidance while grounding concrete preference values in the receiving user’s own memory.

1 INTRODUCTION

  • Pooling histories can expose reusable procedures but also transfer instructions that conflict with the receiving user’s requirements, creating an experience-reachability bottleneck.The central distinction is between access to useful procedures and applicability of user-specific values.Source: §1 · pp. 1–2
  • Across web navigation, online personalized interaction, and multi-session coding, ShareMem improves over matched user-local configurations across four backbone models, especially when relevant local experience is sparse.Further analyses examine consolidation and active preference retrieval, and identify source quality and cross-user preference interference as limits to useful transfer.Source: §1 · p. 2
  • ShareMem combines two-stage experience consolidation and joint local–shared retrieval under a shared entry budget with a preference channel bound to the receiving user.The channel supports initial and agent-initiated recall, while shared guidance indicates which personal constraints to consult and the user’s memory supplies their concrete values.Source: §1 · p. 2

2 RELATED WORK

Prior work develops user-personalized memory, reusable experience abstraction, and cross-agent sharing. ShareMem builds on these strands by consolidating shared experience and grounding its application in the receiving user’s preferences.

Individual agent memory and personalization

  • Prior systems integrate experience with reflection and retrieval or develop long-term storage, while personalized memory grounds assistance in user histories; ShareMem extends this grounding beyond local experience.Examples include Generative Agents, CoALA, MemGPT, MemoryBank, Mem0, A-Mem, and personalized memory systems.Source: §2 · p. 2

Abstracting and maintaining experience

  • ShareMem refines local experience before integrating accepted edits into a shared pool, addressing competition for limited retrieval slots as contributions accumulate.Related work extracts feedback and lessons, develops reusable skills and workflows, and explores experience abstraction, refinement, and maintenance.Source: §2 · p. 2

Sharing experience across users and agents

  • ShareMem combines scope-first experience retrieval with recall from the receiving user’s memory, separating transferable guidance from personal constraints in cross-user sharing.Prior sharing work studies aggregated agent experience, multi-user access control, collective skill evolution, and knowledge applicability across environments or clients; ShareMem also examines preference interference.Source: §2 · p. 2

3 SHAREMEM: SELECTIVE MEMORY SHARING

ShareMem shares reusable procedural experience across users while keeping personal preferences user-bound. It refines experience locally before consolidating accepted edits into shared memory, then retrieves experience and preferences through separate channels.

3.3 JOINT EXPERIENCE RETRIEVAL AND ACTIVE PREFERENCE RETRIEVAL

  • Scope-first retrieval jointly selects local and shared experiences under one entry budget, while a separate user-bound channel lets the agent initially or actively retrieve preferences.Closer-matching scope groups take priority, and lower-priority groups only fill remaining slots; experience guides preference queries but does not trigger them automatically, and the environment fixes the user identity.Source: Figure 1 · p. 3, §3.3 · p. 4

3.1 MEMORY STRUCTURE

  • ShareMem gives each user a dedicated agent and preference store, while local and shared experience collections provide reusable procedures organized by applicability scope.An environment-specific adapter assigns scopes and retrieval priorities; agents resolve personal requirements using the receiving user’s own preferences, not other users’ stores.Source: §3.1 · p. 3

3.2 TWO-STAGE EXPERIENCE CONSOLIDATION

  • Two-stage consolidation first updates local experience using relevant neighbors, then applies accepted local edits as evidence for independently decided shared-memory edits.The managers abstract source-specific values into semantic slots and preference-retrieval guidance; without accepted local edits, shared memory remains unchanged.Source: §3.2 · pp. 3–4

4 EXPERIMENTS

ShareMem is evaluated on web interaction, personalized online tasks, and multi-session coding, comparing shared experience with user-local memory and benchmark-specific reference systems. Experiments also test consolidation, retrieval, local-coverage, evidence-quality, admission, and preference-interference effects.

4.1 EXPERIMENTAL SETUP

  • The experiments use task-specific metrics and controlled memory configurations across Mind2Web, VitaBench 2.0, and MemoryCode.Four backbone models use BGE-M3 embeddings and BGE-Reranker-v2-M3; Mind2Web and MemoryCode use fixed offline experience collections, while VitaBench updates collections online from completed full-reward subtasks.Source: §4.1 · pp. 4–5

4.2 MAIN RESULTS

  • ShareMem beats AWM by 2.43 percentage points in Mind2Web step success, but trails history rewriting and full-context conditioning on VitaBench and scores 3.21 D-MS points below Full-Pref on MemoryCode.It exceeds RAG on VitaBench and substantially outperforms raw-history conditioning on MemoryCode; Full-Pref supplies the complete extracted preference collection.Source: Figure 2 · p. 6, §4.2 · pp. 6–7

4.3 ABLATION STUDIES

  • Two-stage consolidation produces a smaller shared pool and higher D-MS on all three subsets, while append-only storage creates larger pools and lower D-MS.Two-stage writing consolidates local experience before sharing accepted edits and reduces shared-memory induction tokens. Append-only extraction neither revises nor removes entries; under a fixed five-entry retrieval budget, larger pools may increase competition among near-duplicates.Source: Table 4 · p. 7, Figure 3 · p. 7, §4.3 · p. 7
  • Scope-first retrieval achieves higher Mind2Web step success than flat retrieval across tested budgets and retains substantially higher same-site coverage.The comparison uses frozen pools, the same similarity function, and matched retrieval budgets; same-site coverage declines for both strategies as the budget grows.Source: §4.3 · p. 7, Figure 4 · p. 8
  • On MemoryCode, ShareMem improves mean D-MS over PREF by 9.48 points with active preference retrieval, versus 1.37 points with initial recall alone.This supports complementarity between experience guidance and user-specific constraints, but measures combined local and shared experience rather than sharing alone; disabling active retrieval also changes information access and interaction opportunities.Source: Figure 4 · p. 8, §4.3 · p. 8

4.4 FURTHER ANALYSIS

  • Mind2Web sharing gains decline from 4.10 to 1.30 to −0.13 step-success points as same-site local coverage rises from 0.00% to 49.60% to 98.81%.Replacing gold demonstrations with cached agent rollouts reduces ShareMem step success by 5.90 points versus 1.53 for LOCAL. In cold-start online Mind2Web, stricter admission thresholds slow accumulation; as τ rises from 0.5 to 0.999, LOCAL loses approximately 2.6 step-success points, while ShareMem declines from its intermediate-threshold peak, and online scores remain below offline results.Source: §4.4 · pp. 8–9, Figure 5 · p. 9, Figure 6 · p. 9
  • Removing SHARE-ALL’s seven reserved current-user preference slots lowers current-user entries to fewer than two on average and reduces performance, while reservation does not prevent cross-user preference misuse.Audited failures show misuse despite source labels, explicit user identification, and instructions against applying other users’ values, motivating preference retrieval restricted to the receiving user’s own store.Source: Figure 7 · p. 9, Figure 8 · p. 9, §4.4 · pp. 9–10

5 CONCLUSION

ShareMem enables cross-user reuse of experience while grounding its application in each receiving user’s preferences. The findings support sharing reusable guidance without assuming that larger pools or broader preference access improve performance.

5 CONCLUSION

  • ShareMem adds value beyond user-local memory, particularly when relevant experience is unavailable locally.Effective transfer depends on experience quality, consolidation, and access to personal constraints during execution; larger pools or broader access to other users’ preferences do not necessarily improve performance.Source: §5 · p. 10

AI USE STATEMENT

  • GPT-5, GPT-6, and Codex assisted with language polishing, structural reorganization, and related-work retrieval, but not experiments or scientific findings.Authors independently verified all AI-assisted outputs and take responsibility for the final work.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

  • Real-world deployment should require appropriate consent and safeguards against unintended cross-user information sharing.The experiments use public benchmarks, and ShareMem separates reusable experiences from user-specific preferences.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • The paper documents its memory procedures, evaluation setup, benchmark adaptations, ablation protocols, construction costs, and interference audits.Appendix C provides key prompts for experience extraction, memory updates, and preference-aware execution.Source: REPRODUCIBILITY STATEMENT · p. 10

A IMPLEMENTATION AND EVALUATION PROTOCOLS

The appendix describes the benchmark adaptations and evaluation protocols, comparison memory configurations, and procedures for memory construction and retrieval.

A IMPLEMENTATION AND EVALUATION PROTOCOLS

A.1 BENCHMARK PROTOCOLS AND AGGREGATION … B SUPPLEMENTARY ABLATIONS AND MECHANISM EVIDENCE

The supplementary protocols define how experiences and preferences are constructed, accessed, and scored across three benchmarks. Controlled configurations isolate local and shared experience access from user-bound or cross-user preference access, while scope-first retrieval and separate update channels govern application.

A.1 BENCHMARK PROTOCOLS AND AGGREGATION

  • Mind2Web and MemoryCode use fixed offline experience collections, whereas VitaBench 2.0 interleaves experience updates with evaluation.VitaBench accepts experience updates only from fully rewarded subtasks; preference acquisition from user interactions is separate, and each trial starts a new memory trajectory.Source: §A.1 · pp. 12–13
  • Mind2Web evaluates 252 tasks on recorded webpage states rather than live-browser rollouts; VitaBench 2.0 retains 56 users and 771 temporally ordered subtasks; MemoryCode uses three fixed 20-dialogue subsets.MemoryCode subsets contain 395, 413, and 420 coding tasks, respectively; conditions within each subset use the same dialogues and evaluation queries.Source: §A.1 · p. 13
  • Mind2Web scores average task-level element accuracy, action F1, and step success across three runs; VitaBench reports Avg@3, Pass@3, and Pass3 across trials, and MemoryCode’s D-MS equally averages subset scores.Mind2Web main scores retain skipped-step records as zeros; MemoryCode averages outputs within dialogues, then equally across dialogues on a 0–100 scale.Source: §A.1 · p. 13

A.2 MEMORY CONFIGURATIONS

  • The controlled configurations distinguish no memory, preference-only access, local experiences, shared experiences, and cross-user preferences; Mind2Web omits preference conditions, and comparisons match backbone, tasks, and retrieval limits.SHAREMEM keeps preference retrieval bound to the current user and jointly budgets local and shared experiences at K = 5; SHARE-ALL additionally permits source-labeled, non-authoritative other-user preference hints, with at least seven of up to 10 slots reserved for the current user when candidates permit.Source: §A.2 · pp. 13–14

A.3 EXPERIENCE UPDATES AND RETRIEVAL

  • Experience scope follows benchmark structure: Mind2Web prioritizes website, subdomain, and domain; VitaBench uses task domain, and MemoryCode uses one coding scope.Scope matching prioritizes candidate selection but does not guarantee that every experience in a matched collection applies.Source: §A.3 · p. 14, Table 6 · p. 14
  • Local managers update user collections from demonstrations, accepted subtasks, or historical sessions, then shared managers independently consolidate accepted local changes against shared entries.Each manager can add, replace, remove, or leave entries unchanged; local edits are not replayed verbatim into shared memory.Source: §A.3 · pp. 14–15
  • Scope-first retrieval gathers up to 25 candidates per accessible collection, then selects up to five under a joint budget without local/shared quotas within a scope-priority group.Lower-priority groups are used only if capacity remains; Mind2Web and VitaBench apply scoped selection, while MemoryCode jointly reranks within its single coding scope.Source: Table 6 · p. 14, §A.3 · p. 15
  • Preference stores are separate from experiences, and available preference channels support initial retrieval of up to 10 entries plus agent-issued read-only queries during execution.VitaBench preferences update from supplied interactions, while MemoryCode uses a fixed evaluation snapshot; retrieving an experience does not automatically trigger a preference query.Source: §A.3 · p. 15

B.1 TWO-STAGE AND EDITABLE CONSOLIDATION

Editable two-stage consolidation outperforms alternative write paths across ShareMem subsets, while append-only pools show greater lexical redundancy. Two-stage writing also reduces induction cost.

B.1 TWO-STAGE AND EDITABLE CONSOLIDATION

  • Table 4 favors editable two-stage writing over both alternatives on every ShareMem subset, and Table 7 reports lower task-weighted scores under both alternative write paths.Direct writing affects configurations that access shared experiences; append-only also affects LOCAL.Source: §B.1 · p. 15, Table 7 · p. 16
  • Append-only enlarges shared pools and increases entries per lexical cluster in every subset, indicating accumulation of lexically overlapping experiences.The analysis links entries when their word-set Jaccard similarity is at least 0.5, after removing words present in at least 90% of pool entries.Source: §B.1 · pp. 15–16, Table 8 · p. 16
  • Two-stage writing reduces shared-stage token use by a factor of 5.72 and total induction tokens by approximately 49% versus append-only.Local induction costs approximately 2.99M tokens under either path; the comparison uses Qwen-tokenizer construction-stage counts.Source: §B.1 · p. 16, Table 9 · p. 16

B.2 EXPERIENCE RETRIEVAL AND BUDGET

On Mind2Web, shared experience access improves over Local, while the additional benefit of scope priority varies by backbone and retrieval budget. In the Qwen budget sweep, five entries form a compact operating point, and scope-first concentrates retrieved guidance on the evaluation website.

Experience budget

  • In a Qwen seed-42 budget sweep, both strategies improve from one to five experiences then plateau, with scope-first 0.03 to 0.88 percentage points higher at every tested budget.Scope-first also retrieves a larger proportion of same-site entries. Because the budget bounds grounding demonstrations as well as retrieved experiences, five entries are supported as a compact operating point for this setting.Source: §B.2 · p. 16

Retrieval strategy

  • At K = 5, both shared configurations beat Local on Qwen and Gemma; scope-first exceeds flat on Qwen but performs similarly on Gemma.Source: §B.2 · p. 16, Table 10 · p. 17

B.3 COVERAGE AND ADMISSION QUALITY

ShareMem’s incremental value depends on whether recipients already have relevant local experience, while stricter online admission constrains experience supply and coverage.

Local coverage

  • As same-site local coverage rises, ShareMem’s gain shrinks and scores converge near complete coverage; the ordering holds across all three seeds.The experiment reallocates a fixed set of demonstrations across 20 users and measures coverage from retrieval records.Source: §B.3 · p. 17, Table 12 · p. 17

Online admission

  • Stricter admission reduces contributing trajectories and website coverage; at high thresholds, ShareMem retains broader shared-pool coverage and a narrower score range as local performance declines.Unlike the fixed-demonstration assignment experiment, this online sweep combines admission, pool coverage, and subsequent agent interactions.Source: §B.3 · p. 17

B.4 EVIDENCE SOURCE AT FIXED TRAINING-TASK COVERAGE · B.5 ACTIVE PREFERENCE RETRIEVAL

The evidence source affects how much cross-user sharing helps, while reusable experience guidance and active access to user-specific preferences provide complementary benefits. The analyses also qualify these findings by the quality and scope of the evidence and interventions tested.

B.4 EVIDENCE SOURCE AT FIXED TRAINING-TASK COVERAGE

  • The evidence-source intervention retains all 1,009 training tasks and exclusive data assignment, varying whether each task uses a gold demonstration or a cached no-memory rollout.Both experience construction and attached demonstrations use the selected source, and the resulting pools are frozen for evaluation.Source: §B.4 · p. 17
  • The retrieval audit contrasts Local’s same-user experiences from other websites with ShareMem’s site-matched shared experiences, so source quality affects different guidance in each configuration.The analysis motivates balancing evidence quality against the availability of relevant experience.Source: §B.4 · p. 18

B.5 ACTIVE PREFERENCE RETRIEVAL

  • The preference-access comparison uses matched Qwen MemoryCode tasks and fixed preference stores and experience pools, with the same task-query top-10 initial preference recall in all configurations.PREF disables experience guidance; ShareMem enables local and shared experiences. Initial-recall-only variants respond in one turn, whereas active variants can retrieve additional preferences during execution.Source: §B.5 · p. 18

B.4 EVIDENCE SOURCE AT FIXED TRAINING-TASK COVERAGE

  • Replacing gold demonstrations with agent rollouts reduces ShareMem’s advantage, while Local remains comparatively stable when at least half the tasks use gold demonstrations.The intervention jointly changes induced experience content and attached demonstrations, so it measures sensitivity to the evidence source as a whole.Source: §B.4 · pp. 17–18, Table 14 · p. 18

B.5 ACTIVE PREFERENCE RETRIEVAL

  • Experience guidance and active preference retrieval are complementary, with each providing a larger gain when the other is available.The larger experience gain under active retrieval holds in all three subsets; the intervention changes both available information and the opportunity for further interaction.Source: Table 15 · p. 18, §B.5 · p. 18

B.6 CROSS-USER PREFERENCE INTERFERENCE · C PROMPTS

The audit finds model- and benchmark-specific cross-user preference misuse despite source labels and instructions. The accompanying prompts describe how to abstract reusable experiences while keeping concrete preference values grounded in the current user’s memory.

B.6 CROSS-USER PREFERENCE INTERFERENCE

  • The audit tests SHARE-ALL outputs across four backbones and three seeds, requiring traceable foreign values that conflict with recipient preferences and violate failed constraints.Qwen3.6-27B screens cases at temperature zero, then makes three fresh judgments at temperature 0.7; confirmation uses unanimity or a two-of-three majority, and percentages use all exposed failures as the denominator.Source: §B.6 · p. 19
  • Gemma has the highest confirmed misuse rate on MemoryCode, while Gemini and DeepSeek have the highest rates on VitaBench 2.0.The table separates screened candidates from repeat-confirmed cases, particularly for MemoryCode; these rates cover confirmed misuse among exposed failures, while screening misses remain unmeasured.Source: Table 16 · p. 19, §B.6 · p. 19
  • Four unanimously confirmed cases show coding outputs adopting foreign naming conventions and service outputs applying foreign choices despite recipient-specific requirements.The cases link recipient requirements, injected foreign preferences, and recorded outputs; each foreign entry was source-labeled as a non-authoritative hint.Source: §B.6 · pp. 19–20, Table 17 · p. 20

C PROMPTS

  • The prompts section presents key VitaBench 2.0 prompts for memory construction and personalized execution.Source: C PROMPTS · p. 21

C.1 LOCAL EXPERIENCE UPDATES

  • Local-update prompts keep concrete preference values out of experience entries and add, replace, remove, or leave entries unchanged based on reusable procedural evidence.They favor focused task triggers, abstract preference dimensions, bounded recovery, and current task evidence over stored experience guidance.Source: §C.1 · p. 21

C.2 SHARED EXPERIENCE UPDATES

  • Shared-update prompts consolidate reusable cross-user procedures while removing source-user values and preserving distinct task families.Current instructions, current-user preferences, profile, context, and tool observations override shared experience guidance.Source: §C.2 · p. 22

C.3 EXPERIENCE EXTRACTION

  • Experience extraction prompts derive procedures primarily from successful trajectories, using historical interactions only to identify relevant preference dimensions, never their concrete values.Entries use recognizable task triggers, focused retrieval for unresolved aspects, verification, and bounded recovery rather than universal domain guidance or unbounded retries.Source: §C.3 · p. 23

C.4 PREFERENCE-AWARE EXECUTION

  • Preference-aware execution resolves each required personalized field from current-user sources and searches separately for unresolved aspects before acting.Auto-recall may be incomplete; current-user preferences are authoritative, while other-user shared preferences are only weak hints and their exact values must not be copied into actions.Source: §C.4 · p. 24

Related papers

Loading 2609.32511v1…