Mnem.OS

Carry useful context forward.

AI workflows repeatedly lose useful context between sessions. Simply saving every message can mix observations with trusted facts and make later decisions difficult to explain.

What I worked on.

• Scoped memory objects and context packets with source references, confidence, exclusions and audit traces.
• Agent lifecycle events and review-required memory proposals that remain separate from accepted durable knowledge.
• Operator review, contradiction visibility and recovery snapshots alongside MCP and REST integration surfaces.

The implementation includes bounded SQLite-backed search, context packing, agent continuity records, review bundles and recovery routes. Tests exercise sensitivity exclusions, scope isolation, review-required proposals and recovery behavior.