AuthorsZouying Cao, Jiaji Deng, Li Yu et al.
TL;DR
ReMe uses multi-faceted experience distillation plus utility-based refinement to turn procedural memory into a self-evolving pool, letting Qwen3-8B + ReMe (dynamic) reach 55.03% Avg@4 vs 46.20% without memory (+8.83 points).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Procedural memory in many LLM agents follows a “passive accumulation” paradigm, where experiences are stored as static, append-only archives without refinement.
These static systems often mix valid insights with toxic noise, causing agents to misgeneralize in new tasks and degrading performance over time in complex agentic environments.
HOW IT WORKS
ReMe centers on experience acquisition, experience reuse, and experience refinement, combining multi-faceted distillation, scenario-aware indexing, and utility-based deletion into one closed-loop procedural memory lifecycle.
You can think of ReMe like a brain that not only remembers skills but also rewrites and prunes them, more like a constantly edited playbook than a static log file.
This design lets ReMe maintain a compact, high-utility experience pool that adapts to shifting tasks, enabling behaviors a plain context window or raw trajectory store cannot support.
DIAGRAM
This diagram shows how ReMe retrieves, reranks, rewrites, and applies experiences during experience reuse for a single task query.
DIAGRAM
This diagram shows how ReMe is evaluated on BFCL-V3 and AppWorld, including fixed vs dynamic settings and key ablations.
PROCESS
Experience Acquisition
ReMe samples multiple execution trajectories with LLMexecute, then uses multi-faceted distillation in experience acquisition to build structured experiences with scenarios, content, keywords, confidence, and tools.
Experience Reuse
Given a new task query, ReMe performs experience reuse by retrieving top K experiences via usage scenario indexing, optional reranking, and adaptive rewriting into task specific guidance.
Experience Refinement
After task execution, ReMe runs experience refinement with selective addition from successful trajectories, failure aware reflection for retries, and utility based deletion using retrieval and success statistics.
Experience driven Inference
Throughout the lifecycle, ReMe enables experience driven inference where LLMexecute reasons with distilled experiences, progressively evolving its procedural memory across BFCL V3 and AppWorld tasks.
KEY CONTRIBUTIONS
ReMe framework for agent evolution
ReMe integrates experience acquisition, experience reuse, and experience refinement into a closed loop, enabling agents to autonomously distill and maintain high quality procedural experiences across tasks.
reme.library procedural memory dataset
ReMe introduces reme.library, a fine grained procedural memory dataset with structured success patterns and failure lessons distilled from diverse agentic tasks for studying procedural memory.
Memory scaling effect across Qwen3 models
ReMe shows that Qwen3-8B + ReMe (dynamic) reaches 55.03% Pass@4 vs 54.65% for Qwen3-14B without memory, and Qwen3-14B + ReMe (dynamic) surpasses Qwen3-32B without memory on Avg@4 and Pass@4.
RESULTS
Avg@4
34.94% Avg@4
+7.29 over No Memory (Qwen3-8B average Avg@4 27.65%)
Pass@4
55.03% Pass@4
+8.83 over No Memory (Qwen3-8B average Pass@4 46.20%)
BFCL-V3 Avg@4
45.17% Avg@4
+4.84 over No Memory on BFCL-V3 with Qwen3-8B (40.33%)
AppWorld Pass@4
42.06% Pass@4
+9.21 over No Memory on AppWorld with Qwen3-8B (32.85%)
ReMe is evaluated on BFCL-V3 and AppWorld, both tool-augmented multi-turn benchmarks testing complex agentic tasks. The gains in Avg@4 and Pass@4 show that ReMe’s self-evolving procedural memory substantially improves success probability over No Memory baselines.
BENCHMARK
ReMe is evaluated on BFCL-V3 and AppWorld, both tool-augmented multi-turn benchmarks testing complex agentic tasks. The gains in Avg@4 and Pass@4 show that ReMe’s self-evolving procedural memory substantially improves success probability over No Memory baselines.
BENCHMARK
Avg@4 on BFCL-V3 for Qwen3-8B with different memory systems.
BENCHMARK
Avg@4 for Qwen3-8B with trajectory-level vs keypoint-level experience acquisition in ReMe (fixed).
KEY INSIGHT
ReMe enables Qwen3-8B + ReMe (dynamic) to reach 55.03% Pass@4, slightly above Qwen3-14B without memory at 54.65% Pass@4.
This is surprising because larger Qwen3-14B is expected to dominate, but ReMe shows that high quality procedural memory can offset model size advantages.
WHY IT MATTERS
ReMe unlocks self-evolving procedural memory where agents continuously distill, adapt, and prune experiences instead of hoarding static trajectories.
Builders can now deploy smaller backbones with ReMe to achieve or surpass larger No Memory systems, making long horizon, tool-augmented agents more practical under tight compute budgets.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…