History search

Search everything your agents have already done on a project — across Claude Code, Codex, Antigravity and Vibe — and reopen the original conversation from a result.

The Memory panel’s Search tab answers one question: what happened on this project before? It searches the transcripts of every supported CLI, not just the one in front of you.

What gets indexed

OFM ingests conversations directly from each CLI’s own store — it never keeps a second copy of your transcripts. Supported: Claude Code, Codex, Antigravity and Mistral Vibe. Each ingested fragment is typed, so results are readable at a glance:

  • Your message — what you asked
  • File edited — what the agent changed
  • Command / output — what it ran and what came back
  • Error — what failed

Indexing is incremental. The status line reports what happened, for example Indexed: 12/48 sessions, 3 400 fragments — the ones already ingested are skipped.

Searching

Type in the search box. Full-text search runs immediately, with no model and no download.

Semantic search is optional. Click Encode to build embeddings on top of the index; the first run downloads an on-device model of roughly 90 MB. After that, encoding runs locally and the status shows how many fragments remain. Use it when you remember the idea but not the words — full-text search is enough when you remember the command.

Reopening a conversation

A result is not a dead quote. Click a Claude Code result and OFM reopens the original conversation in a live session tab, positioned on that project. This is the payoff of the session model: transcripts stay in the CLI’s store, so OFM can hand the id straight back to claude --resume.

For CLIs that cannot be resumed by id, the result is copied instead, and the status line says so explicitly rather than silently doing nothing.

Where the data lives

One SQLite store per project, under project-memory/ in the config directory. No network, no upload, no API call — see Project memory.

From history to knowledge

Raw history is a haystack. The point of searching it is usually to recover a decision — and a decision you had to search for twice belongs in project knowledge, where agents read it automatically. The Distill latest session button does exactly that conversion.

Frequently asked

Which CLIs have their history indexed?
Claude Code, Codex, Antigravity and Mistral Vibe. OpenCode, Kimi Code and Cursor Agent do not expose a transcript format OFM can ingest today.
Is semantic search required?
No. Full-text search works immediately with no download. Semantic encoding is optional and pulls a roughly 90 MB on-device model the first time you enable it.

Last updated