kernex-memory
Persistent memory store for the Kernex AI agent runtime.
Backed by SQLite (via sqlx), the store records:
- Conversation history and FTS5-indexed message recall
- User facts and project-scoped facts
- Lessons (reward-weighted recall), outcomes, and consolidator state
- Token usage and per-dimension breakdown (input / output / prompt-cache reads / prompt-cache creations)
- Phase checkpoints for multi-stage workflows
- An audit log of permission decisions
The schema is migration-managed; downgrading is not supported. Files are created with mode 0o600 on Unix, and the parent directory is locked to 0o700.
You usually consume memory through kernex-runtime, which wires Store into the request pipeline (history → facts → recall → outcome storage). Depend on kernex-memory directly when building a side-channel tool that reads or migrates memory state.
Documentation
- API reference: https://docs.rs/kernex-memory
- Project overview: https://github.com/kernex-dev/kernex
License
Apache-2.0 OR MIT.