MCP Tools Reference

Eidet exposes 13 tools via the Model Context Protocol. These are the tools AI agents call directly.

Transports

Transport	Command	Use case
stdio	eidet mcp	Local AI clients (Claude Code, Cursor)
HTTP	POST http://localhost:19380/mcp	Network MCP clients

Tool Listing

eidet_store

Store a new memory.

{
  "repo": "P:\\MyProject",
  "content": "The UserService caches results in Redis with a 5-minute TTL",
  "type": "observation",
  "tags": ["redis", "caching", "user-service"],
  "importance": 0.6,
  "source": "claude-session",
  "session_id": "sess-abc123",
  "supersedes": null
}

Returns: { "id": "memories/...", "success": true } or rejection reason.

eidet_recall

Recall memories by semantic query (hybrid vector + full-text search).

{
  "repo": "P:\\MyProject",
  "query": "how does caching work",
  "limit": 10,
  "type": "insight"
}

Returns: Array of scored results with content, type, tags, importance, staleness warnings.

eidet_context

Get the compact L0+L1 session context. This is typically the first tool an agent calls.

{
  "repo": "P:\\MyProject"
}

Returns: A text block under 600 tokens with memory counts and top-scored memories.

On first call for a new repo (when no memories exist), automatically triggers intake to scan project files.

eidet_forget

Soft-delete a memory with an optional reason.

{
  "id": "memories/P--MyProject/observation/abc123",
  "reason": "This information is outdated"
}

Creates an audit trail observation recording what was forgotten and why.

eidet_feedback

Report whether a recalled memory was useful.

{
  "memory_id": "memories/P--MyProject/insight/abc123",
  "was_used": true
}

• was_used: true — Echo: boosts importance and confidence
• was_used: false — Fizzle: reduces importance and confidence

eidet_history

View the version chain for a memory (current + all ancestors via supersession).

{
  "id": "memories/P--MyProject/insight/abc123"
}

eidet_intake

Ingest project files (CLAUDE.md, README.md, .editorconfig, package config) as seed memories. Splits by headings, deduplicates by content hash, extracts entities.

{
  "repo": "P:\\MyProject"
}

eidet_link

Create a cross-repo link between two memories.

{
  "source_id": "memories/P--ProjectA/insight/abc",
  "target_id": "memories/P--ProjectB/insight/def",
  "relation": "depends-on"
}

eidet_consolidate

Group related observations and create insights. Runs FadeMem decay on all memories.

{
  "repo": "P:\\MyProject"
}

eidet_maintenance

Run the full 7-stage maintenance pipeline:

1. TTL expiry
2. Observation retention (default 90 days)
3. Deduplication sweep (Jaccard similarity 0.85)
4. Importance decay
5. Orphan cleanup
6. Backfill enrichment (entities + one-liners)
7. Auto-consolidation

{
  "repo": "P:\\MyProject"
}

eidet_export

Export all memories as formatted markdown.

{
  "repo": "P:\\MyProject"
}

eidet_pack_export

Export memories as a portable .eidet pack file for sharing.

{
  "repo": "P:\\MyProject"
}

eidet_pack_import

Import a .eidet pack and optionally mount it as a layer.

{
  "repo": "P:\\MyProject",
  "data": "...",
  "mount_as_layer": true
}

Agent Integration Pattern

A typical AI agent session flow:

1. Session starts
2. Agent calls eidet_context → gets L0+L1 summary
3. Agent works on task, calls eidet_recall as needed
4. Agent discovers something → calls eidet_store
5. Agent uses a recalled memory → calls eidet_feedback(was_used: true)
6. Agent finds a memory wrong → calls eidet_forget with reason
7. Session ends (memories persist for next session)

CLAUDE.md Instructions

Use eidet instructions to generate ready-made instructions for your CLAUDE.md:

# Print to stdout
eidet instructions --print

# Install into ~/.claude/CLAUDE.md
eidet instructions --install

# Create in project root
eidet instructions --project