Memory

Search first: /mnemonic:search {relevant_keywords} Capture after: /mnemonic:capture {namespace} "{title}"

Run /mnemonic:list --namespaces to see available namespaces from loaded ontologies.

Mnemonic Core

Memory system operations. See references/ for full documentation.

Trigger Phrases

Capture Triggers (silently capture when user says):

Decisions (namespace: _semantic/decisions):

• "I've decided", "let's use", "we're going with"
• "we'll use", "I'm choosing", "going forward with"
• "the decision is", "we chose", "decided on"

Learnings (namespace: _semantic/knowledge):

• "I learned", "turns out", "discovered that"
• "TIL", "gotcha", "found out", "realized"
• "the root cause was", "it was because"

Patterns (namespace: _procedural/patterns):

• "always use", "never do", "when X do Y"
• "the pattern is", "we should always", "convention is"
• "standard approach", "best practice here"

Blockers (namespace: _episodic/sessions/blockers):

• "I'm stuck on", "blocked by", "can't figure out"
• "hitting a wall", "the problem is", "struggling with"

Context (namespace: _semantic/knowledge):

• "important to know", "keep in mind", "context is"
• "background", "constraint", "requirement"

Recall Triggers (search memories when user asks):

• "what did we decide about", "how do we handle"
• "what's our approach to", "check for decisions"
• "any learnings about", "patterns for"
• "remind me", "what do we know about"
• "search memories", "recall", "find memories"

Silent Capture Protocol

When to capture: Only capture when the user explicitly states a decision, learning, pattern, or blocker in their message. Do NOT capture based on Claude's own suggestions or recommendations.

When trigger phrases are detected:

1. Capture to appropriate namespace WITHOUT announcing
2. Do NOT say "I'm capturing this" or similar
3. Continue natural conversation flow
4. Memory creation is invisible to user

Do NOT capture:

• Claude's own suggestions (only user-stated facts)
• Hypothetical scenarios ("we could use...")
• Questions or exploratory discussion

Visibility & Review

Users can review captured memories at any time:

# Check session status and recent captures
/mnemonic:status

# List recent memories (last 7 days)
find ${MNEMONIC_ROOT} -name "*.memory.md" -mtime -7 -exec basename {} \;

# Search for specific captures
/mnemonic:search {keywords}

If a user asks "what did you capture?" or "show me my memories":

• Run /mnemonic:status to show memory stats
• Use search to find recent relevant memories
• This is NOT a violation of silent capture - user explicitly requested visibility

Capture

/mnemonic:capture {namespace} "{title}"

Namespaces: _semantic/decisions | _semantic/knowledge | _procedural/patterns | _episodic/blockers | _episodic/sessions

MANDATORY: Before creating any memory, check for duplicates:

1. Search for existing memories with similar keywords:

   rg -i "{key_words}" ${MNEMONIC_ROOT}/ --glob "*.memory.md" -l | head -5

2. If matches found, read the top result
3. If it covers the same topic → UPDATE existing memory (use Edit tool) instead of creating new
4. If related but different → create new memory with a relates-to relationship
5. Only create brand new if no matches found

Recall

rg -i "{keywords}" ${MNEMONIC_ROOT}/ --glob "*.memory.md" -l

Minimal Memory Format

CRITICAL: Generate real values. NEVER write placeholders like "PLACEHOLDER_UUID" or "PLACEHOLDER_DATE".

Generate a UUID and timestamp FIRST, then write the file:

# Generate these BEFORE writing the memory file
UUID=$(uuidgen 2>/dev/null | tr '[:upper:]' '[:lower:]')
[ -z "$UUID" ] && UUID=$(python3 -c "import uuid; print(uuid.uuid4())")
DATE=$(date -u +"%Y-%m-%dT%H:%M:%SZ")

Type must match namespace path:

• _semantic/* → type: semantic
• _procedural/* → type: procedural
• _episodic/* → type: episodic

Then create the file with the real generated values:

---
id: <the real UUID generated above>
title: "Your actual title here"
type: semantic
namespace: _semantic/decisions
created: <the real DATE generated above>
temporal:
  validFrom: <the real DATE generated above>
  recordedAt: <the real DATE generated above>
  lastAccessed: <the real DATE generated above>
  decay:
    model: exponential
    halfLife: P90D
    strength: 1.0
provenance:
  sourceType: agent_inferred   # user_explicit if the user stated it directly
  agent: claude
  confidence: 0.9
---

# Your actual title here

Content here.

Half-life defaults:

• _semantic/decisions: P180D (long-lived)
• _semantic/knowledge: P90D
• _procedural/patterns: P180D (long-lived)
• _episodic/*: P30D (fast decay)

References

• references/capture.md - Full capture workflow
• references/recall.md - Search and retrieval patterns
• references/schema.md - Complete MIF Level 3 schema
• references/examples.md - Working examples