swafra — Persistent Memory Skill

You have persistent memory across sessions via swafra. Use it proactively — never wait for the user to ask.

Commands

Run these via bash. All output is JSON.

Retrieve memory (MANDATORY — call before your first response)

python3 -m swafra.engine_cli get_context "USER_TOPIC_OR_QUESTION"

Store knowledge (call WITHOUT being asked)

python3 -m swafra.engine_cli add "TEXT_TO_STORE" "descriptive-title"

Search memory

python3 -m swafra.engine_cli search "query" 8

List all stored sources

python3 -m swafra.engine_cli list

Delete a source

python3 -m swafra.engine_cli delete "SOURCE_ID"

Graph walk (explore connections from a chunk)

python3 -m swafra.engine_cli walk "CHUNK_ID" 2 10

Rules

1. ALWAYS call get_context before your first response — use the user's question/topic as the query
2. ALWAYS call add when the user shares ANY of:
   • Their name, role, preferences, or personal context
   • Project decisions, architecture, or technical choices
   • Corrections to your behavior or rules to follow
   • Documents, meeting notes, or long-form content
3. Never say "I don't have context from previous sessions" without calling get_context first
4. Use descriptive titles when storing: preference-editor, project-swafra, meeting-2026-07-28
5. Err on the side of storing too much — if it has future value, store it

How it works

• Data is stored locally in ~/.scimap/ as JSON (no cloud, no database)
• Text is chunked using Leiden community detection, embedded locally, and graph-linked
• Retrieval uses hybrid scoring: BM25 + vector cosine + entity overlap + n-gram matching
• Facts are tracked with lifecycle management (old facts get superseded, stale chunks penalized)

If python3 doesn't work

Try python3.12 or python3.11 instead.