Digest is not an insert: information is decomposed and linked, and the maturity model decides what survives. That is the answer to memory stores that grow until retrieval degrades — decay is deliberate, and hub nodes are protected from it. Raw data is never deleted; what changes is prominence, and a recall restores it immediately. Note also what running it costs nothing: keyword search and storage work with no API key at all.
A memory layer that sits between your AI tools and your notes rather than replacing either. Conversations and note content become nodes and links in a single graph, so retrieval can follow a decision back to the discussion behind it and forward to what it changed, instead of ranking a flat list of facts.
- `brain_prepare` loads memory context — called once at the start of a conversation, and its result indexes what to ask for next
- `brain_recall` navigates and retrieves, following links between memories rather than only matching keywords, and each recall reinforces what it returns
- `brain_digest` stores new information, decomposing it and weaving it into the existing network
- Nodes carry four maturity dimensions — activity, refinement, connectivity, independence — that decide whether a memory is reinforced, connected, crystallised or forgotten
- Background metabolism: daily decay of inactive memories with highly connected hubs protected, and weekly divergent scanning that looks for node pairs sharing neighbours but lacking a direct link
- Note sources: Logseq with file watching and incremental sync, Obsidian vault import including Canvas and wikilinks, and read-only Apple Notes sync
- A desktop client for browsing the graph, watching the metabolism and tuning parameters
Node.js >= 18 is the only hard requirement — it runs out of the box with BM25 text search and local storage. Everything lives in one SQLite file on your machine, exportable to Markdown. The macOS desktop app connects your tools from a settings panel; from source it is `npm install`, `npm run build`, `npm start`. Optional: an LLM provider (Anthropic, Google Vertex AI or Gemini) enables automatic extraction, reconsolidation, divergent scanning and crystal emergence, and an embedding provider (Ollama, Vertex AI or Gemini) adds semantic search — without them those steps are skipped rather than failing. The npm package is `tidemind`, version 0.2.89.