Definition
Plain language
A memory system that lets an AI agent default to written notes and only turn a lesson into reusable code once it's proven worth it.
As stated in the literature
A self-evolving agent-memory framework that stores experience as text (plans, facts, pitfalls) and crystallizes plans into callable tools only after a recurrence gate, codifying from queries rather than raw trajectories.
Why it matters: It avoids the waste of prematurely hardening every experience into code, keeping memory cheap while still building durable tools for what actually recurs.
For example, an agent jots down a useful troubleshooting plan as a note and only turns it into reusable code once it has needed that same plan several times.
Heard on the show
“… The paper is called Metis, out of the Chinese University of Hong Kong and Huawei, and it does two separable things — a diagnostic …”Episode 168 — When Turning Experience Into Code Makes Your AI Agent Dumber