"A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E." — Tom Mitchell
We study how autonomous agents build persistent, environment-grounded memory. Our goal is to move beyond stateless tool use and naive retrieval by enabling agents to construct structured models of the environments they operate in, distill successful experience into reusable procedures, verify those procedures, and maintain them over time through versioning, correction, and retirement. We are building a general framework that can be applied across domains and environements.
We do this by exploring the following directions:
- Improved harness designs for improved learning
- Dedicated memory for environments
- Evaluations and benchmarks for testing solutions