MCP Tools

When you connect memex to Claude, it gets access to 13 tools. Claude uses these automatically, you don't have to ask.

Tools

search_notes parameters

Example: "What did we decide about auth last April?" → Claude calls search_notes with query="auth decision", date_from="2025-04-01", date_to="2025-04-30".

save_note behavior

save_note requires a layer parameter, past, state, or rule. The classification rules are baked into the tool description so Claude picks correctly: past for records of what happened (default when in doubt), state for current plans, rule only when you explicitly ask.

Before saving, Claude is instructed to:

1. Call list_tags and list_folders to see existing taxonomy
2. Pick from existing tags/folders rather than inventing new ones
3. Use [[Title]] wiki-link syntax for references to people, projects, or concepts

If save_note returns a ⚠️ similar notes warning, Claude switches to update_note on the most relevant existing note instead of creating a duplicate.

The response also includes any flashbacks, see Flashback in tool responses below.

update_note layer guards

update_note is layer-aware:

• past → refused with error: 'PAST_IMMUTABLE' and a suggestion to save a new [Amendment] note that wiki-links back to the original. Claude is instructed to follow the suggestion automatically.
• rule → refused with error: 'RULE_USER_ONLY'. Claude surfaces its proposed change in chat instead of writing it.
• state → updates normally.

This keeps your audit trail intact (past records stay frozen) and prevents Claude from silently rewriting its own behaviour rules.

House Rules auto-inject

Notes with layer = 'rule' are appended to the MCP server's instructions on boot, under a ## House Rules section. Claude sees them at the start of every conversation, no search_notes call required. Coding style guides, personal preferences, and policy notes belong here.

Changes are picked up on the next Claude Desktop / Claude Code restart.

Flashback in tool responses

Both save_note and search_notes append a 🔗 Flashback section when older notes from a different folder match the current context:

Saved note #142: "Auth approach revisited"

🔗 Flashback, older notes from a different context:
- 124 days ago: #48 "Auth Architecture Decision" (82% match)

For search_notes, the flashback anchor is the top result. Claude is instructed to surface these to you when relevant, they're how "I wrote about this 3 months ago in a different project" connections actually become visible.

Tunable via env:

get_note + backlinks

get_note returns the full note content and a "Referenced by" section listing every note that links to this one via [[Title]] syntax. Backlinks are resolved at save/update time.

Inference engine tools

These expose the inference engine to Claude. Detection and invalidation stay deterministic in memex; the only LLM step is synthesizing an inference's summary, and Claude does that, never memex.

• get_signals(type?, status?), lists un-synthesized patterns. Defaults to new and re-runs detection (dirty-flagged, so it's free on a static corpus). Use it when you ask "what threads are in my notes", "what haven't I synthesized", or "what should I write about".
• update_signal_status(id, status), dismiss noise so it never resurfaces, or snooze it. Claude confirms with you before dismissing.
• list_inferences(status?) / get_inference(id), read stored hypotheses. Staleness is re-checked first, so a changed source shows up as stale.
• mint_inference(...), persist a hypothesis.

Two guardrails make this safe to trust:

1. Inferences are pull-only, never injected. They're excluded from search_notes and never auto-loaded into Claude's context. Claude pulls them on purpose and cites them as hypotheses ([Inference #42, conf 0.78, from notes #1, #2]) never as facts you "know". This prevents the brain from echoing its own past guesses back as truth.
2. No auto-mint. mint_inference requires confirmed: true, which Claude only sets after you explicitly approve. It presents the hypothesis in chat first; minting in the background is refused at the code level, not just by instruction.

On connect, the server greeting includes a startup count of new signals and stale inferences (a nudge to look), but never their content.

Claude's automatic behavior

Claude is configured to:

• Always search before answering any question that might relate to past conversations, decisions, or projects
• Save at end of conversation when technical decisions were made, key insights emerged, or you explicitly said "remember this"
• Prefer update over save when new content belongs with an existing note