LifeMem: Enabling Lifelong Experience Reuse for LLM Agents
Yuli Qiu, Yutong Li, Wei Su, Zeming Liu, Wanxiang Che, Heyan Huang, Haifeng Wang, Yuang Guo
cs.CL
Quick verdict
LifeMem clusters trajectories into workflows to distill reusable agent skills; broad 10-env study supports gains, but no significance tests or variance given.
Summary
LifeMem is a lifelong learning framework for LLM agents that enables experience reuse across heterogeneous environments. During learning, it clusters accumulated interaction trajectories by underlying workflow (using a directional-statistics MRL constraint to split bloated clusters) and distills environment-invariant skills from each cluster with an LLM; at inference the agent retrieves same-environment trajectories plus their clusters' skills. Evaluation spans 10 environments in 5 scenario types (12,489 training / 1,230 test tasks), including 2,310 newly generated trajectories (1,541 successful) for four environments lacking training data, with GPT-4o-mini, DeepSeek-v3.2-exp, and Qwen3-32b backbones. LifeMem attains the best average overall performance (44.13/48.02/46.52) versus Synapse, ExpeL, and AutoSkill, and is the only method with consistently positive backward transfer (+5.66 to +6.68), indicating reduced forgetting; ablations show both trajectories and skills contribute, and enhanced retrieval alone does not close the gap. Zero-shot transfer to FEVER and τ-bench-airline also favors LifeMem. Main limitations: no significance tests or repeated-run variance despite small test sets and temperature 0.3; headline results use a curated similarity-ordered stream under which the random-stream advantage nearly vanishes; LifeMem trails the memory-free ReAct baseline on some datasets; and closely related workflow-induction systems (AWM, Hermes) are cited but not compared.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Clustering trajectories by underlying workflow, with an MRL-based diversity constraint controlling splits, prevents cross-environment retrieval interference in a shared agent memory.
- Distilling environment-invariant skills from each cluster and pairing them with environment-aligned trajectories yields the best lifelong performance (average overall performance 44.13/48.02/46.52 across the three backbones, Table 2).
- LifeMem is the only evaluated method with positive average backward transfer (+5.66 to +6.68), meaning subsequent tasks reinforce rather than degrade earlier ones.
- Distilled skills enable zero-shot transfer to unseen task formulations and environments (FEVER +22.5%, τ-bench-airline +18.8% over ReAct on GPT-4o-mini, Table 4).
- Task streaming order materially affects performance; grouping structurally similar tasks or using a small buffer recovers much of the loss under uncurated streams (Table 5).
Strengths
Unusually broad evaluation scope
10 environments across 5 scenario types with 12,489 train / 1,230 test tasks (Table 1), three backbones (GPT-4o-mini, DeepSeek-v3.2-exp, Qwen3-32b), plus 2,310 newly generated trajectories (1,541 successful, §5.1) for four environments lacking training data — a genuine data contribution.
Ablations support the two-component design
Table 3: trajectories-only 42.84, skills-only 37.61, joint 44.13 vs ReAct 28.84 (GPT-4o-mini); LifeMem with gpt-4o-mini-distilled skills (43.18) even beats vanilla gpt-4.1 ReAct (40.42), indicating gains stem from memory design rather than a stronger LLM.
Positive backward transfer directly evidences less forgetting
Retrieval-upgrade control isolates the architectural claim
Human-verified memory structure at modest cost
Weaknesses
No statistical rigor
Favorable streaming in headline results
Unanalyzed losses to the memory-free baseline
Missing nearest-neighbor baselines
Closed-model dependency and thin human-eval reporting
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
LifeMem (arXiv 2609.12655, cs.CL) targets inter-environment lifelong learning for LLM agents: one agent accumulates interaction trajectories from 10 heterogeneous environments and must reuse them without cross-environment interference or catastrophic forgetting. The central claims are: (1) clustering trajectories by underlying workflow (rather than surface semantic similarity) prevents retrieval interference; (2) distilling environment-invariant skills per cluster, paired with environment-aligned trajectories, yields the best overall performance and positive backward transfer; (3) skills enable zero-shot transfer to unseen formulations and environments; (4) task-stream order materially affects results. The paper also contributes 2,310 newly generated trajectories (1,541 successful) for four environments lacking training data.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is LifeMem?
LifeMem is a lifelong learning framework for LLM agents (arXiv 2609.12655). It stores interaction trajectories from many environments in memory clusters organized by underlying workflow, distills environment-invariant skills from each cluster with an LLM, and at test time gives the agent both same-environment example trajectories and cross-environment skills, reducing forgetting and improving transfer.
How much does LifeMem improve over prior memory methods?
Across 10 environments and three LLM backbones, LifeMem's average overall performance is 44.13 (GPT-4o-mini), 48.02 (DeepSeek-v3.2-exp) and 46.52 (Qwen3-32b), beating Synapse, ExpeL and AutoSkill. It is the only tested method with positive average backward transfer (+5.66 to +6.68), meaning later tasks tend to help rather than hurt earlier ones. Note the paper reports single runs without significance tests.
Is LifeMem's code and data available?
The paper states the dataset and code are available at github.com/BITHLP/LifeMem, including over 2k newly generated interaction trajectories for four environments (τ-bench, HotpotQA, LifelongAgentBench, ToolQA) that previously lacked training data. Building the memory uses closed OpenAI models (GPT-4.1, GPT-5-mini), which affects exact reproducibility.
Does LifeMem still work if tasks arrive in random order?
Performance drops from 44.13 to 39.63 (GPT-4o-mini) under a fully random task stream, roughly matching the Synapse baseline (39.18). A simple buffer that caches 5 random trajectories per task before committing them recovers much of the loss (41.64), which the authors present as evidence the method remains usable on uncurated streams.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-6 are embedded images; only captions and in-text summaries were extractable, so the transfer-event distribution (Fig. 3) and cluster-size/cluster-type statistics (Figs. 5-6) could not be independently verified against the stated numbers.
- The provided metadata lists the last author as 'Yuang Guo' while the paper text prints 'Yuhang Guo' (corresponding author); this review follows the paper text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.