← News

Agents & MCPSources: arXiv · 18 Nov 2026

EvoDisco lets agent memory learn which update to make, not just when

A new arXiv paper frames memory evolution in LLM agents as evolutionary strategy discovery, with a learnable policy choosing among four memory operators and improving results on three agent benchmarks.

In plain words

AI agents that act over many tasks keep notes about past experience, called memory. The EvoDisco paper proposes letting the agent learn how to reorganize that memory itself, instead of following fixed rewrite rules. A learned policy picks one of four memory operations and is trained with a two-level scoring signal.

What was announced

The authors argue that existing online-memory methods fix both the operators that transform memory and the rules selecting them. EvoDisco instead treats memory evolution as a discovery problem: a learnable policy, conditioned on the current memory state and recent task history, selects and instantiates one of four operators. Abstract compresses related memories, Refine decomposes overloaded entries, Bridge connects weakly linked experiences, and Reframe reorganizes retrieved reasoning chains.

Training uses a bi-level fitness signal. An inner signal scores each transformation by retrieval-quality improvement, regularized by memory complexity, while an outer signal scores strategic efficacy over a window of past decisions using operator-utility regret, utility-prediction error, and selection diversity. Under a one-pass online plus frozen-transfer protocol, the paper reports gains over MemP and MemRL on ALFWorld, WebShop, and LifelongAgentBench (OS and DB), with the largest improvements on ALFWorld: +8.8 online and +12.9 frozen-transfer points over MemP, and +16.4 and +17.9 over MemRL. Ablations credit both fitness levels and the operator library.

What it does not show

  • Evidence is limited to the paper's abstract; the full experimental setup, baselines' configurations, and robustness of the reported gains cannot be verified from this excerpt.
  • Reported improvements are the authors' own measurements on three benchmarks and have not been independently confirmed.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.