evolve

To extract governance rules from agent memory, the tool analyzes LanceDB reflections and LEARNINGS.md corrections to group similar insights into candidate policies. These policies are classified as MUST, NEVER, or ALWAYS rules for human review.

Converting reflection memories into AGENTS.md rules involves clustering recurring patterns from your vector database, deduplicating similar insights, and outputting structured candidate rules. This process ensures continuous improvement of model behavior and governance.

Yes, generating candidate rules requires LanceDB to query and retrieve stored reflection memories. The tool depends on LanceDB alongside your LEARNINGS.md corrections to extract recurring patterns and classify them into actionable policies.

To deduplicate and cluster similar rules from agent learnings, the tool groups recurring insights extracted from your memory database and correction logs. It generates a traceable changelog for version control and rollback while outputting unique candidate policies.

Yes, you can use an optional learnings_extractor to expand rule coverage from memory reflections. It extracts extended rules from your correction records to improve pattern discovery and increase the overall coverage of generated governance policies.