memory-store

Content: [What happened, what was tried, what worked/failed, key lesson]

Tags: #role #topic #success|#failure

**Step 2: Extract metadata**
Parse the formatted text to extract:
- `title`: Plain text without markdown (from "**Title:**")
- `preview`: Plain text without markdown (from "**Preview:**")

**Step 3: Detect role for storage**
Determine which role collection to store in based on task context. Use role_mapping below. Default to "OTHER" if unclear.

**Step 4: Search for duplicates**
Use `search_memory` with full formatted text as query, `roles=["detected_role", "OTHER"]`, `limit=10`.

Note: The `roles` parameter tells MCP which collections to search.

**Step 4.5: Review and fetch specific memories**
After search returns previews:
1. LLM reads the preview texts (10 results)
2. Select 3-4 that might match or conflict with new memory
3. FETCH those specific ones using `batch_get_memories(doc_ids, roles=["detected_role", "OTHER"])`
4. Read fetched full content to decide: duplicate? merge? update?

**Important:** Don't rely on similarity score - rely on LLM review of actual content.

**Step 5: Decide action**
Based on fetched content review:
- Near-identical exists → MERGE (combine, delete old)
- Related exists → UPDATE (enhance existing)
- Pattern emerges from 2+ episodic → GENERALIZE to semantic
- Different topic → CREATE new

**Step 6: Store**
Use `store_memory(document, role, metadata)` with:
- `document`: Full formatted markdown text (Title + Preview + Content + Tags)
- `role`: Collection name (backend, frontend, scrum-master, qa, OTHER)
- `metadata`: 3-field dict extracted from document:
```json
{
  "title": "Plain text title (extracted from **Title:** line)",
  "preview": "2-3 sentence summary (extracted from **Preview:** line)",
  "content": "[Full formatted markdown document - same as document parameter]"
}

API Clarification:

• document = the full markdown text
• role = which collection to store in
• metadata = {title, preview, content} - 3 fields extracted from document