Jason Weston's self-improving memory lifts agent test scores from 36.1 to 56.0

jaseweston · x · 2026-11-17

Jason Weston shares a thread on Self-Improving Memory: the agent runs LOAD before each subtask and OFFLOAD after, with files carrying what was learned. In the Discover step, a meta agent reads trajectories, finds where memory use breaks, and rewrites the LOAD/OFFLOAD prompts; best candidates (5 runs each) become parents for iteration. Optimized harness reaches an avg test score of 56.0 vs 44.6 for the best hand-written version and 36.1 for the generic one. The meta (discover) agent and task agent are the same model (Kimi-K2.6), so the model learns memory skills it discovered itself.

Related event: Meta's Self-Improving Memory Lets Agents Discover and Distill Memory Skills(4 posts)→

Original post →

More from coding & agent

coding & agent channel →