Codex Session History

Local search across past Codex sessions with incremental indexing and an MCP interface.

Published: 09/24/26, 3:30 PM Updated: 09/24/26, 3:35 PM

Put together with AI using materials from my project and my edits.

AI-translated from Russian. Read the original.

Codex Session History helps recover a decision, fact, or unfinished task from earlier Codex sessions. It builds a local search index over JSONL files and exposes bounded results through MCP.

Search across working history

Queries can filter by text, project, status, and date range. Four MCP tools divide the regular workflow: index_status reports index state, refresh_sessions picks up changes, search_sessions finds relevant sessions, and get_session returns a selected excerpt.

Each refresh reparses only changed files. This makes the index suitable for regular Codex work: earlier context is retrieved on demand as selected excerpts.

Local data boundary

JSONL files are mounted read-only inside the container. The derived index stays in PostgreSQL within a Compose volume, while MCP communicates over stdio with no published network ports. Raw transcripts never enter the repository or become part of a public artifact.

The current version uses PostgreSQL full-text search. The schema already includes a pgvector-ready table for future measured reranking.