LLM agents can improve across tasks by writing their own execution experience to external memory and reusing it in later decisions, without updating model parameters. The central design question is whether to keep only observed trajectories and outcomes, or also to retain the rules, summaries and skills inferred from them. Keeping abstractions lets agents reuse strategies directly, avoiding repeated derivation from raw trajectories. However, if stored abstractions contain errors, those errors propagate through retrieval and reuse, shape execution trajectories, and become the basis for new abstractions, creating an error‑reproduction feedback loop: abstraction → execution → re‑abstraction. To decouple experience accumulation from abstraction persistence we propose the Anti‑Persistence memory design. It continuously gathers trajectories and outcomes while reconstructing task‑specific guidance on demand. An adaptive selector uses prior outcomes to decide whether and how to synthesize guidance, which is discarded after execution, preventing it from being stored as a directly retrievable object. Indirect feedback via retained trajectories remains possible, but direct reuse of faulty abstractions is blocked. Experiments across various benchmarks and backbones show that Anti‑Persistence achieves the highest overall held‑out success rate and the lowest average LLM‑call cost among all memory designs. These results support separating experience accumulation from abstraction persistence.
Review