Memory and learning
Private fallible context separated from shared Fleet truth and reviewed improvement.
Private Mate memory helps one persistent Agent apply earlier feedback and stable context. It is fallible context, not Fleet authority, current-state proof or permission.
For example, memory may say the Captain prefers decisions to reach a particular company surface. That preference helps shape communication, but current PostgreSQL rows decide what remains unresolved, provider state decides what changed and current instructions decide what is allowed.
Memory boundaries
- Memory belongs to one Mate’s persistent home and is private by default.
- The Captain or supervising Mate can ask it to remember, correct, forget, pin, pause or resume.
- A Mate may propose a useful learning upward; it does not edit another Mate’s PVC.
- Conflicts are surfaced rather than silently rewriting a newer instruction.
- A stale-write guard prevents one session from overwriting a newer memory revision.
Organization-wide learning follows a different path: freeze evaluation evidence, identify the smallest reproducible cause, make one reviewed improvement and rerun the held-out behavior. A good memory hint may inspire that review, but it does not bypass it.
Last updated on