Cognitive Context Manager
A local-first Codex plugin that turns long-running agent activity into compact working context, typed memory, open loops, artifact state, and measurable retrieval quality.
01 / The problem
What needs to change
Long AI coding sessions can lose decisions, failed attempts, file state, and unresolved intentions when context becomes crowded or is compacted into a vague handoff.
02 / The approach
A focused path through the complexity
The plugin continuously consolidates meaningful events into typed local memories, then retrieves a bounded context brief instead of replaying raw transcripts into the active prompt.
03 / Key capabilities
What the system is designed to do
- Typed episodic, semantic, procedural, open-loop, artifact, and safety memory
- Compact working-context and project-state retrieval
- Secret redaction, conflict handling, hygiene, and quarantine
- Checkpoint, handoff, and context-dividend reporting
- Local dashboard and adaptive guidance with explicit review controls
04 / Architecture
Technology and system shape
- Codex plugin manifest, lifecycle hooks, MCP server, skill, and CLI
- SQLite persistence with full-text and hybrid semantic retrieval
- Cloud embeddings through existing auth with deterministic local fallback
- Encrypted optional file sync and local-only dashboard
- Effectiveness benchmarks and trace explainability
05 / Results
What the repository demonstrates
- The public beta supports explicit memory, compact context retrieval, privacy controls, effectiveness reporting, and a local operational dashboard.
- The design keeps current files and project instructions authoritative while treating memory as supporting context.
06 / Lessons
What the work teaches
- A better memory layer should reduce context noise, not simply store more conversation.
- Retrieval quality is only credible when the system can explain and measure what it brought back.
07 / Screenshots
Approved visuals
No public screenshots yet
Add only approved screenshots to the project record.