A personal project for searching previous Codex conversations through MCP. It captures session history in the background and returns excerpts with citations to the original conversation. I built it so my agent can refer to earlier work across sessions.
The implementation uses Python and SQLite, with keyword search and optional embeddings. Capture and keyword retrieval work independently of the embedding model. Citations identify immutable revisions; filters scope results by project, session, role, and time. It runs locally or as an authenticated shared service. In shared mode, each device captures its own sessions and queues uploads to one primary; agents search the combined history. This syncs searchable evidence, not native Codex session lists.
For remote use, keep the service behind Tailscale or an equivalent private network; do not expose it directly to the public internet. Private networking does not replace HTTPS or Session Search device authentication.
| Reference | Contents |
|---|---|
| Architecture | Capture, storage, retrieval, and service boundaries |
| Evaluation | Ranking experiments, results, and limitations |
| Setup | Installation and background capture, carried out by an agent |
| Development | Code layout and checks |
Development version · macOS / Linux · Python 3.11+ · Apache-2.0