Infoguana — AI skill for Claude Code

Iguanas are ectotherms that depend on external heat to function. Infoguana — *info* + *iguana* — gives LLM agents the same kind of external lifeline: a cross-project memory with typed-graph notes, hybrid retrieval, and MCP integration. FastAPI + SQLite (with sqlite-vec + FTS5) + HTMX UI + MCP. Captures notes from phone/laptop, classifies them through the Claude CLI or any OpenAI-compatible endpoint, and serves them to any MCP-capable agent — Claude Code and Codex both ship with an installer — so every agent, in every repository, is backed by the same shared memory.

How it works

Infoguana is a typed graph of short notes with hybrid retrieval, designed to feed LLM agents only the slice of memory relevant to the current turn.

**Notes have types and tags.** Every note is one of `memory`, `feedback`, `rule`, `skill`, `reference`, `plan`, `task`, `feature`, or `idea` — the type changes how it's surfaced (e.g. `feedback` is sticky guidance the agent should re-read, `plan` and `task` enter a lifecycle, `reference` is a pointer). Tags are agent-curated at write time; `tag_suggest` ranks existing vocabulary first so tags don't drift into singletons.

**Retrieval is hybrid and budgeted.** `search` fuses BM25 (FTS5) and cosine similarity (sqlite-vec) over a single ranking. Hits come back as **previews** — haiku-sized 1–5 line summaries generated at write time — so an agent can triage 20 results for a few hundred tokens and only pull full bodies (`get` / `get_many` / `expand_top=N`) for the ones worth quoting.

**SessionStart loads a layered, token-budgeted context pack.** On a new session — Claude Code or Codex, both ship a hook — the agent's first turn is packed with: a **skill manifest** (one line per available skill — name, id, and trigger condition — exempt from the token budget, with bodies fetched by id on demand), **global-scope rules** (cross-project guidance the agent