More memory, worse agents: Anti-Persistence discards synthesized guidance to cut errors
IUB · hf · 2026-10-21
A paper from Indiana University and Tencent HY identifies a hidden failure mode in LLM agent memory: errors in stored abstractions (rules, summaries, skills) propagate through retrieval and reuse, and can shape future trajectories that regenerate the original errors — a cycle the authors call error reproduction, extending errors' persistence beyond direct reuse of faulty knowledge.
Their Anti-Persistence memory design:
- Accumulates trajectories and outcomes but does not persist guidance synthesized from them
- An adaptive selector reconstructs task-specific guidance on demand and discards it after execution, preventing direct reuse as a stored object (retained trajectories still provide indirect feedback)
- Across benchmarks and backbones, Anti-Persistence achieves higher held-out success and the lowest average LLM-call cost among evaluated memory designs
Takeaway: separate experience accumulation from abstraction persistence — store trajectories, not conclusions.
More from coding & agent
- Pinekin to open-source: batch-sync conversation context across AI agent harnesses — yangyi · 2026-10-21
- CoEvo-RSI: NUS-led paper tackles brain-body mismatch in self-improving agents — Jiaqi Su · 2026-10-21
- Open-source Hermes video toolchain turns agent requests into finished MP4s — Teknium · 2026-10-21
- Zed tested 25 MCP tool search approaches, picked weighted BM25 over LLMs and embeddings — jobergum · 2026-10-21
- Argon excels as a video director, says Bindu Reddy, orchestrating SeeDance and ElevenLabs — bindureddy · 2026-10-21
- OmniGraph Open-Sourced: Git-Style Branches and Merges for AI Agent Memory — techNmak · 2026-10-21