memories-learn

name	memories-learn
description	Extract learnings from session transcripts and store in semantic memory database
allowed-tools	Task, Bash(simba *)

Check the dispatch mode:

simba config get hooks.learn_async

Resolve the transcript for THIS project (never the global latest.json — it is a single symlink overwritten by whichever session compacted last, across all projects, so it cross-wires sessions):

simba transcript pending --json

This prints the newest pending_extraction transcript whose project_path matches the current working directory: {transcript_path, session_id, project_path} (or {} + exit 1 if there is nothing to extract for this project — in that case stop, there is no work to do). Use those three values below.

Build this Task prompt:

Read the transcript at <TRANSCRIPT_PATH> and extract learnings to store in the semantic memory database.

For each learning found, store it by running:
simba memory store --type <TYPE> --content "<LEARNING>" --context "<CONTEXT>" --confidence <SCORE> --session-source "<SESSION_ID>" --project-path "<PROJECT_PATH>"

LEARNING TYPES:
- WORKING_SOLUTION: Commands, code, or approaches that worked
- GOTCHA: Traps, counterintuitive behaviors, "watch out for this"
- PATTERN: Recurring architectural decisions or workflows
- DECISION: Explicit design choices with reasoning
- FAILURE: What didn't work and why
- PREFERENCE: User's stated preferences

RULES:
- Be specific - include actual commands, paths, error messages
- Confidence 0.95+ for explicitly confirmed, 0.85+ for strong evidence
- Skip generic programming knowledge Claude already knows
- Focus on user-specific infrastructure, preferences, workflows
- Keep content within the configured `memory.max_content_length` (default 200 characters), use context field for details
- Preserve proper nouns, file paths, and identifiers verbatim — never replace them with generic words
- Preserve numeric precision: keep exact values exact; never weaken an exact number to a range or approximation
- Resolve relative dates to absolute ones (e.g. "yesterday" -> the actual date)

Extract 5-15 quality learnings.

Dispatch using the Task tool with subagent_type=memory-extractor:

• If hooks.learn_async is "true": set run_in_background=true (fire and forget)
• Otherwise: dispatch normally and wait for completion

After the extractor finishes (synchronous mode only), mark the transcript done so it isn't re-extracted on the next run:

simba transcript mark-extracted <SESSION_ID>

(In async mode, skip this — the background agent owns completion.)