Research map / Multi-agent systems and games / LLM agents
Remember the Decision, Not the Description: A Rate-Distortion Framework for Agent Memory
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Frames agent memory as a decision-centric rate-distortion problem and proposes DeMem, an online K-slot memory learner that splits slots only on certified decision conflict, with regret bounds and LoCoMo/LongMemEval gains.
Problem
Budgeted runtime memory for long-horizon LLM agents is typically organized by descriptive criteria (relevance, salience, summary fidelity), which do not capture what histories must remain distinguishable to support good decisions under a fixed memory budget.
Approach
Measure memory quality by decision distortion (suboptimality gap of a shared action) rather than description; derive an exact forgetting boundary (histories can share a state iff they admit a common epsilon-optimal action), covering/packing bounds, and an online learner that refines a K-slot partition only when lower-confidence certificates show decision conflict.
Key results
- On LoCoMo with GPT-4o-mini, DeMem reaches 0.911 overall vs. 0.888 for Mnemis and 0.681 for FullContext.
- On LoCoMo with GPT-4.1-mini, DeMem reaches 0.920 overall vs. 0.906 for Mnemis.
- Description-based retrieval recovers 66% of gold evidence under matched budget vs. 83% for DeMem.
Surprising findings
- Descriptive similarity is only a statistically significant but weak predictor of evidence compatibility (Spearman rho = 0.103, AUC = 0.548) on LoCoMo.
- Certified splits fire on only 4.6% of LoCoMo routing events.
Limitations
- Pairwise cannot-link certificates do not compose into cluster compatibility; practical method uses greedy relaxation of an NP-hard partition problem.
- Real-benchmark instantiation uses a candidate-restricted answer-conflict test rather than the exact theoretical certificate.
- i.i.d. assumption over answer-time decision instances is a modeling simplification for dialogue histories.
Experimental setup
Datasets: LoCoMo LongMemEval MemoryArena Decoupled Bandit (synthetic)
Benchmarks: LoCoMo LongMemEval MemoryArena Decoupled Bandit
Metrics: overall accuracy cumulative regret memory-distortion tradeoff success rate
Baselines: FullContext RAG LangMem Mem0 Zep Nemori EMem-G Mnemis CLUB Feature-KMeans
Builds on: UCB CLUB bisimulation metrics Hoeffding concentration
Closest papers by meaning
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.91)
- EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.91)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.90)
- Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks (ICML 2026 · similarity 0.90)
- Memory-T1: Reinforcement Learning for Temporal Reasoning in Multi-session Agents (ICLR 2026 · similarity 0.90)
- AdaMEM: Test-Time Adaptive Memory for Language Agents (ICML 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.