Definition
Plain language
An approach where an AI agent learns mid-conversation by editing a small slice of its own weights, not just by taking notes it looks up later.
As stated in the literature
A self-evolving agent memory that periodically distills session context into QA flashcards and trains them into a small writable LoRA adapter (fast weights), with reinforcement learning making the agent better at writing memory it can learn from.
Why it matters: It moves agents from merely storing and looking up notes toward actually learning from their experiences during use, while keeping the changes small and contained.
For example, after a long support chat, the agent turns key facts into flashcards and trains a small piece of itself on them so it genuinely remembers next time.
Heard on the show
“TMEM is just what happens when you let that parametric channel be nonzero and, more importantly, trainable.”Episode 114 — Agents That Rewrite Their Own Weights Instead of Just Taking Notes