Project record
Hermes Learn Policy
Policy controls for safe, task-aware agent memory and skill learning.
Show · explain · prove
A concise first record.
Hermes Learn Policy adds task-aware guidance before model calls and checks proposed learning tool calls before Hermes runs them.
Hermes keeps sole ownership of validation, locking, persistence, history, and rollback. The plugin adds no model-callable tools or external service.
My contribution
The parts I owned.
- Designed the foreground, autonomous-review, and Curator policy lanes
- Implemented pre-call and post-call hooks around native learning tools
- Added bounded read evidence and one same-file retry rule
- Kept profile isolation and native Hermes ownership explicit