engram

name	engram
description	Use Engram MCP tools to save and semantically retrieve thoughts from the user's local knowledge base. Trigger on: "remember this", "save this", "note that", "add to my brain", "what do I know about X", "search my notes", "find my thoughts on Y", "do I have anything about Z", "show recent thoughts", "what did I save". Always search Engram before answering from memory when the user references their own notes, ideas, or previously saved content.
license	MIT
metadata	{"version":"v1","publisher":"engram"}

Engram — Local Knowledge Base

You have access to an Engram MCP server that stores and semantically searches the user's personal knowledge base. Everything runs locally — no cloud, no API costs.

When to Use Engram

Trigger phrase	Action
"remember this", "save this", "note that", "add to my brain"	save_thought
"what do I know about X", "find my thoughts on Y", "do I have anything about Z"	search_engram
"search my notes", "look in my brain", "check my knowledge base"	search_engram
"show recent thoughts", "what did I save lately", "my last N entries"	list_recent_thoughts
"delete thought ID N", "remove that thought"	delete_thought
"how many thoughts", "brain stats", "knowledge base stats"	engram_stats

Proactive rule: if the user asks a question and references their own notes/ideas/memory, always call search_engram first before answering from your own training data.

---

Tools Reference

search_engram

Semantic similarity search — finds thoughts by meaning, not keywords.

{
  "query": "ideas about developer productivity",
  "limit": 10,
  "threshold": 0.3
}

• query (required) — natural language description of what you're looking for. Longer, descriptive queries work better than single keywords.
• limit — max results (1–50, default 10)
• threshold — minimum similarity score (0.0–1.0, default 0.3). Lower = more results but less relevant. Raise to 0.6+ when you want only high-confidence matches.

Examples:

• User: "what have I thought about AI pricing?" → {"query": "AI pricing cost models"}
• User: "find my notes on React" → {"query": "React frontend development"}
• User: "do I have anything about my sleep?" → {"query": "sleep health habits"}

---

save_thought

Embeds and stores a thought locally. Called when the user wants to capture something.

{
  "content": "The key insight from today's standup: ship small, get feedback fast.",
  "tags": "process,agile"
}

• content (required) — the full text to save. Be specific — more detail = better search recall.
• tags — optional comma-separated tags for filtering (e.g. ai,productivity,ideas)

Examples:

• User: "remember: GraphQL subscriptions cause memory leaks if you forget to unsubscribe" → save with tag graphql,bugs
• User: "note that the Postgres EXPLAIN ANALYZE output showed a seq scan on the users table" → save with tag postgres,performance

---

list_recent_thoughts

Browses entries in reverse chronological order. Use when the user wants to review what they've saved without a specific query.

{
  "limit": 20,
  "offset": 0
}

• limit — entries to return (1–100, default 20)
• offset — pagination offset (default 0)

---

delete_thought

Permanently removes a thought by its numeric ID (shown in search/list results).

{ "id": 42 }

Always confirm the ID with the user before deleting — deletion is permanent.

---

engram_stats

Returns total count, date range, and unique tag count. No parameters needed.

{}

---

Behavior Guidelines

1. Search before answering — when the user says "what did I think about X" or "what do I know about Y", call search_engram first. Do not answer from training data alone.

2. Save verbatim when instructed — when the user says "remember this: [content]", save their exact words. Don't paraphrase unless they ask you to.

3. Confirm saves — after save_thought succeeds, echo the content back briefly so the user knows what was stored and at what ID.

4. Lower threshold when results are sparse — if search_engram returns 0 results, retry with threshold 0.1 before telling the user nothing was found.

5. Suggest saving mid-conversation — if the user shares a decision, insight, or fact that seems worth keeping, proactively offer: "Want me to save this to Engram?"

---

Proactive Knowledge Capture

Save to Engram without being asked when the session reveals something a future session would genuinely benefit from knowing. Judge by whether it would survive in a commit message — if not, it belongs in Engram.

Save proactively when:

• An architectural decision was made with non-obvious rationale (why X over Y)
• A bug's root cause wasn't immediately apparent (the non-obvious thing that caused it)
• A framework, library, or toolchain gotcha was discovered (would hit someone again)
• A multi-step workflow was figured out manually that isn't documented anywhere
• A constraint or invariant in the codebase was uncovered (things that must stay true)

Don't save:

• Routine implementation work with obvious intent
• Things already documented in README, CLAUDE.md, or comments
• Trivial fixes where the cause and solution are both obvious
• Session housekeeping (file reads, searches, listing things)

How to capture:

{
  "content": "Decided to use sqlite-vec over Chroma — at personal-note scale (hundreds of entries) a dedicated vector DB process is overkill; sqlite-vec gives cosine search in the same SQLite file with no extra process.",
  "tags": "architecture,decision"
}

Be specific — vague entries don't retrieve well. Include the why, not just the what.