codex-session-recall

Codex Session Recall is a local-first tool for recovering evidence from earlier Codex conversations. It reads the Codex catalog in read-only mode, parses conversation messages from rollout JSONL files, builds a disposable SQLite FTS5 index, and returns a bounded bundle of matching excerpts with session IDs and source line numbers. The project describes itself as a search-before-context layer for long-running AI work.

The tool separates processing old sessions from using recalled evidence later. It builds a safe, disposable index and lets an agent query it only when prior work may affect the answer. Raw sessions remain the source of truth; nothing is uploaded, and the tool does not write to the Codex state database or the original transcripts.

Multi-term recall is strict first, preferring a single message containing all query terms. If that produces no match, it may fall back to coverage across multiple messages from the same session, while keeping evidence lines separate and marking the result as term_coverage_fallback. Results from different sessions are never merged. An optional assess command exposes a friction trigger for integrations and returns a structured should_recall decision.

Status is 0.2.0-alpha research preview. It requires Python 3.10+ with SQLite FTS5 and a local Codex installation. The runtime has no third-party dependency. Limitations include Codex adapter only, lexical recall rather than semantic/vector search, and no automatic model summarization or memory-palace writeback.