bootstrap-memory

name	bootstrap-memory
description	Seed OpenMemory with existing knowledge from key-prompts, decisions, tool learnings, and behavioral prefs. Safe to re-run — uses per-entry dedup.
args	[]

Bootstrap OpenMemory

Incremental seeding of OpenMemory with project knowledge. Safe to re-run — checks each entry for duplicates before storing.

Steps

0. Setup auto memory symlinks. Run ~/.claude/scripts/setup-memory.sh. This creates the machine-specific projects/<encoded>/memory/ directories and symlinks them to the portable memory/*.md files. If the script fails, warn but continue — OpenMemory seeding doesn't depend on symlinks.

1. Health check. Call openmemory_list(user_id="global", limit=1). If unreachable, abort: "OpenMemory not reachable. Ensure Ollama is running (ollama serve) and the OpenMemory MCP server is registered."

2. Seed key prompt patterns. Read each file in ~/.claude/.claude/tracking/key-prompts/. For each ## entry:

   • Extract the title, category, context, prompt, and why-it-worked fields
   • Compute simhash = md5(compressed_entry)[:16]
   • Call openmemory_query(query=<full compressed entry text>, k=3, user_id="global") — for each result: compare the first 80 chars of content_preview against the entry (case-insensitive, whitespace-normalized); if they match, skip (count as skip_count). If no content match but a result has matching simhash in metadata, also skip.
   • Otherwise: openmemory_store(content="<compressed entry>", tags=["prompt-pattern"], user_id="global", metadata={"sector": "procedural", "simhash": "<simhash>"})
   • Count stored entries as prompt_count

3. Shadow active decisions. Call pm_list_decisions(status="active"). For each decision:

   • Compute simhash = md5(decision_summary)[:16]
   • Call openmemory_query(query=<full decision summary text>, k=3, user_id="proj:dotclaude") — for each result: compare the first 80 chars of content_preview against the entry (case-insensitive, whitespace-normalized); if they match, skip (count as skip_count). If no content match but a result has matching simhash in metadata, also skip.
   • Otherwise: openmemory_store(content="Decision [id]: <title> — <summary>", tags=["decision"], user_id="proj:dotclaude", metadata={"sector": "semantic", "decision_id": "<id>", "simhash": "<simhash>"})
   • Count stored entries as decision_count

4. Seed tool learnings. Read ~/.claude/tool-learnings.md. For each - prefixed line (skip headers and blank lines):

   • Extract the learning text (strip leading - )
   • Compute simhash = md5(learning_text)[:16]
   • Call openmemory_query(query=<learning_text>, k=3, user_id="global") — apply same dedup logic (first 80 chars content match OR simhash match → skip)
   • Otherwise: openmemory_store(content="<learning_text>", tags=["tool-learning"], user_id="global", metadata={"sector": "procedural", "simhash": "<simhash>"})
   • Count stored entries as learning_count

5. Shadow behavioral prefs. Query correction_groups DB for promoted/manual entries:

   • Run: sqlite3 ~/.claude/.claude/epics.db "SELECT text FROM correction_groups WHERE status IN ('promoted') OR source='manual' AND text != '' ORDER BY updated_at DESC;"
   • For each non-empty result:
     • Compute simhash = md5(text)[:16]
     • Call openmemory_query(query=<text>, k=3, user_id="proj:dotclaude") — apply same dedup logic
     • Otherwise: openmemory_store(content="<text>", tags=["behavioral-pref"], user_id="proj:dotclaude", metadata={"sector": "procedural", "simhash": "<simhash>"})
   • Count stored entries as pref_count

6. Report. Print: "Bootstrapped N memories (P prompt patterns, D decisions, L tool learnings, B behavioral prefs). Skipped J duplicates."