Research map / Multi-agent systems and games / LLM agents
The Compliance Trap: Diagnosing How AI Agents Consume Conflicting 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.
Proposes the Entry-Propagation-Recovery (E-P-R) framework for diagnosing how agents consume conflicting memory, and reports a "compliance trap" on WebArena and a new MemTrapBench where similar compliance rates cause larger absolute damage for stronger agents.
Problem
Existing agent memory work focuses on supply (writing, storing, retrieving memory), leaving unclear how models consume retrieved memory across multi-step trajectories, especially when the memory is task-wrong.
Approach
Decomposes memory consumption into Entry, Propagation, and Recovery phases measured through paired no-memory vs. memory-injected trajectories with early/persistent/late injection schedules, and identifies a compliance trap where compliance rate is roughly constant across models but post-compliance success collapses, so stronger models lose more.
Key results
- Persistent conflicting memory lowers Qwen3.5-27B success by 21.2 pp on MemTrapBench (paired DPC +26.6 pp, CI [+19.6, +33.5]).
- Recommendation compliance rate is 63-72% across five WebArena models while post-compliance success falls to 17-31%.
- Retry-on-fail with helpful memory reaches 50.6% on WebArena for Qwen3.5-27B vs 40.3% no-memory and 41.6% always-injected.
Surprising findings
- Compliance rate with conflicting memory is roughly constant across models, so stronger models suffer larger absolute damage.
- A 'verify before acting' warning reduces conflicting damage but also reduces helpful gain, making agents uniformly less compliant.
Limitations
- Evidence limited to web-browsing agents; transfer to terminal or embodied settings untested.
- WebArena 77-task subset selected on memory sensitivity inflates effect sizes.
- Does not propose or train a memory management controller or intervention.
Experimental setup
Datasets: WebArena MemTrapBench
Benchmarks: WebArena MemTrapBench MiniWoB++
Metrics: success rate delta recommendation compliance rate damage per compliance trajectory divergence recovery rate
Baselines: No-memory policy Always persistent helpful injection Retry-without-memory
Builds on: Reflexion ExpeL Agent Workflow Memory WebArena BrowserGym
Closest papers by meaning
- Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks (ICML 2026 · similarity 0.89)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.87)
- G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.87)
- MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.87)
- What Do Agents Learn from Trajectory-SFT: Semantics or Interfaces? (ICML 2026 · similarity 0.87)
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.86)
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.