Stored abstractions can reproduce errors in LLM agent memory
Original title: More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents
Anti-Persistence rebuilds task guidance on demand and discards it, with the lowest average LLM-call cost tested.
The paper studies agents that improve by storing execution experience externally rather than updating parameters. Keeping inferred rules and skills can regenerate errors through abstraction, execution, and re-abstraction. Anti-Persistence retains trajectories and outcomes but synthesizes guidance only for the current task. Across benchmarks and backbones, the authors report higher held-out success than the other memory designs evaluated.
Unconfirmed: Success and cost comparisons are reported by the paper's authors, not independently verified.
- Errors persist beyond direct reuse of the original faulty knowledge.
- An adaptive selector uses prior outcomes to decide how to synthesize guidance.
- Synthesized guidance is discarded after execution and is not stored.
- Indirect feedback through retained trajectories can still occur.
From the original article
1 min read at huggingface.co
LLM agents can improve across tasks without updating model parameters by storing experience from their own executions in external memory and reusing it to guide later decisions. A central design question is whether to retain only observed trajectories and outcomes or also the rules, summaries, and skills inferred from them.
By Ruhan Wang, Kishan Panaganti, Dongruo Zhou · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.