ReMe: A Deep Dive into the Dynamic Procedural Memory Framework for AI Agents

Forum topic · ✨步子哥 · 2026-02-25

Summary

ReMe is a dynamic procedural memory framework developed jointly by Shanghai Jiao Tong University and Alibaba Tongyi Lab, released in December 2025 as an official extension of the AgentScope multi-agent framework. Moving beyond static 'append-only' memory systems, ReMe enables agents to evolve from passive recording to active self-improvement through multi-dimensional distillation of execution traces, context-adaptive experience reuse, and utility-based memory refinement (selective addition plus proactive pruning of stale memories). Experiences are formalized as structured tuples with trigger conditions, content, keywords, confidence scores, and tools. Benchmark results are notable: on BFCL-V3 with Qwen3-8B, ReMe lifted Avg@4 from 0.4033 to 0.4450 (+10.3% relative) and Pass@4 from 0.5955 to 0.6577; on AppWorld it delivered a 13.9% relative Avg@4 gain; and on a tool-memory benchmark scores rose 14.88%. Its most striking finding is a 'memory scaling effect': Qwen3-8B plus ReMe outperforms a memoryless Qwen3-14B, and Qwen3-14B plus ReMe beats a memoryless Qwen3-32B, suggesting high-quality memory can substitute for parameter scaling at a fraction of inference cost. The system identifies an optimal retrieval range of 5-7 experiences. ReMe installs via pip (reme-ai), uses ChromaDB by default with MongoDB support planned, and targets dialogue, tool use, robotics, and automation scenarios.

ReMe: Dynamic Procedural Memory Framework for Agents — Research Deep Dive

ReMe (https://github.com/agentscope-ai/ReMe) is an open-source dynamic procedural memory framework jointly developed by Shanghai Jiao Tong University and Alibaba Tongyi Lab, released December 2025 as an official AgentScope extension. This article is a structured summary of a detailed Chinese research report.

Key points

Paradigm shift: ReMe moves agent memory from "passive recording" to "active evolution" via three mechanisms: multi-dimensional distillation of noisy execution traces, context-adaptive experience reuse, and utility-based refinement (selective addition, failure-aware reflection, and utility-based deletion).
Core definition: A dynamic *procedural* memory framework focused on "how to do" knowledge. Each experience is a five-tuple E = ⟨ω, e, κ, c, τ⟩ (trigger condition, content, keywords, confidence, tools).
Lifecycle management: hierarchical working / short-term / long-term memory; multi-dimensional retrieval (semantic + structural + temporal + value); context-aware recall tuned to output 5–7 best-matching experiences; memory "metabolism" via utility-scored deletion (delete when f(E) ≥ α and u(E)/f(E) ≤ β).
Key modules: Working Memory (smart context compression and attention management), Long-term Memory (layered indexing, dynamic clustering, intelligent archiving), Memory Fusion (conflict detection and strategy synthesis across multiple experiences), and Memory Reflection (metacognitive self-audit of the memory system).
Architecture: three-layer design (interface / core / storage), pluggable embedders (Sentence-BERT, E5, BGE), abstract storage backends (ChromaDB default; MongoDB in progress; Elasticsearch/Milvus under evaluation). Install via pip install reme-ai; minimal config needs only LLM and embedding API keys in .env.

Benchmark results

BFCL-V3 (Qwen3-8B, thinking mode): Avg@4 0.4033 → 0.4450 (+10.3% relative), Pass@4 0.5955 → 0.6577 (+10.4%). Outperforms A-Mem and LangMem baselines; dynamic refinement adds ~0.5% over fixed.
AppWorld (Qwen3-8B, non-thinking): Avg@4 0.1497 → 0.1706 (+13.9% relative); Pass@4 0.3285 → 0.3631 (+10.5%). LangMem degraded (−9.4%) on this benchmark while ReMe improved — evidence of robustness on complex, multi-step tasks.
FrozenLake (100 random maps): pass rate 0.66 → 0.72 (+9.1% relative), driven by both remembered success paths and validated failure lessons.
Tool-memory benchmark (three simulated search tools, Qwen3-30B-Instruct): score 0.672 → 0.772 (+14.88%), with tool-selection accuracy up 18.5%.

Memory scaling effect

The report's most striking finding: smaller models with ReMe outperform larger memoryless models — Qwen3-8B + ReMe (dynamic) beats Qwen3-14B without memory, and Qwen3-14B + ReMe beats Qwen3-32B without memory. Since retrieval overhead is roughly 5–10% of LLM inference cost, this suggests high-quality memory can substitute for parameter scaling, enabling tiered deployments (8B+ReMe for routine tasks, larger models for novel complex cases).

Hyperparameter insight

Sensitivity analysis over K = 1–20 retrieved experiences shows success rises monotonically to K = 5–7 (a performance plateau), then flattens or declines with growing latency beyond K > 7 — echoing cognitive psychology's "7±2" working-memory finding. Default is K = 5.

Application areas

Dialogue systems: unlimited-context multi-turn coherence, user preference personalization, and continuous dialogue-strategy learning.
Tool use / function calling: API usage-pattern learning, error-recovery experience accumulation, multi-step planning templates.
Embodied AI & robotics: skill deposition from physical interaction, cross-scene skill transfer via abstracted representations, adaptive decision-making in dynamic environments.
Emerging areas: autonomous driving scene understanding, AR/VR interaction optimization, and scientific research automation.

Ecosystem and roadmap

Hosted at https://github.com/agentscope-ai/ReMe; current release v0.2.0.6 (Beta), semantic versioning.
Deep AgentScope integration via a "Memory-Aware Agent" base class; YAML/environment-variable-driven configuration; async I/O and multi-level caching for non-blocking performance.
Roadmap: MongoDB support (short term), deeper multimodal memory including video and cross-modal retrieval (mid term), and cloud-native vector database integrations (long term).