Definition
Plain language
A system that lets an AI agent learn from past tasks by keeping a searchable notebook of lessons, while the underlying model stays frozen and swappable.
As stated in the literature
A model-agnostic experience-learning framework storing distilled skills and lessons as nodes in a semantic graph, retrieving via personalized-PageRank diffusion and utility-aware ranking, and training a small copilot to set retrieval parameters using a with-minus-without reward.
Why it matters: It lets agents improve from experience without retraining the underlying model, so accumulated know-how carries over even when the model is swapped out.
For example, after solving a tricky task, the agent saves the lesson in its notebook so it can look it up the next time a similar problem appears, no matter which model is running underneath.
Heard on the show
“… The paper is called "ExpGraph: Model-Agnostic Experience Learning with Graph-Structured Memory for LLM Agents," it went up on …”Episode 106 — Giving Agents a Notebook Instead of New Weights: How ExpGraph Lets Frozen Models Learn