FAQ

Does my code get sent anywhere? No. Anchoring, hashing, indexing, retrieval, and both hooks are entirely local. v0.2 removed the one feature that called an LLM.

Do I need an API key or a vector database? No and no. Retrieval is SQLite FTS5 plus anchor and trigger matching.

Where did extract and inject go? Deleted in v0.2. extract was an LLM pass over session transcripts — write-side sophistication that the evidence says buys 3-8 points where retrieval buys 20, and its fallback path silently saved unanchored, unverifiable memories. inject dumped memories at session start, the point of maximum context dilution; it was never once referenced in our benchmark transcripts.

Where did convention and reference types go? Also deleted. CLAUDE.md and your docs already hold that knowledge, and hosts inject CLAUDE.md for free. Memories written under those types in v0.1 load as decision.

Why does an episode need triggers? Because otherwise it can never be pushed — only searched, which is the weakest channel. The trigger is what lets the memory resurface at the moment the same failure happens again.

Should memories be committed to git? Yes. That is the team-sharing mechanism: your teammate clones the repo and their agent immediately has your failed attempts. The index is gitignored and rebuilds itself.

Why return stale memories at all instead of hiding them? Because the reasoning usually outlives the code. Hiding it loses knowledge; flagging it lets the agent judge.

Does it work with Gemini CLI, Cursor, Codex, or other agents? Partly, and the difference matters. The MCP tools (remember, recall, deprecate) work in any MCP host, Gemini CLI included. The hooks do not - surface and guard are wired through Claude Code's PreToolUse/PostToolUse settings format, which is Claude Code specific.

Since v0.2 the hooks are the primary channel, that is a real limitation rather than a footnote: on non-Claude-Code hosts you get agent-initiated search only, which is the weaker half of the product. Both surface and guard are plain CLI commands that read hook JSON on stdin and print JSON out, so any host with an equivalent hook mechanism can drive them - but we have only tested Claude Code, and we will not claim otherwise until we have.

What languages support symbol anchoring? Python, JavaScript, TypeScript, TSX. Line-range and whole-file anchoring work for any language, and unsupported files degrade gracefully rather than failing.

Does legendary make my agent faster or cheaper? Unproven, and we will not claim it until a valid benchmark says so. Our first attempt was retracted for measuring nothing — see benchmark.

Is it production-ready? v0.2 is a breaking redesign of a young tool. The core — anchoring, staleness, push delivery — is tested and works. Expect the API to keep moving.