Decision Extraction Pipeline

Claustrum stores all Git raw events, always.

Then it runs a two-stage memory pipeline:

1. Raw Event -> Activity Memory (predictable)
2. Raw Event -> Decision Memory (LLM classification)

Flow

Important Rule: Keywords Are Scheduling Only

decision_keyword_policies affects LLM processing priority only.

• It does not create memories by itself.
• It does not confirm decisions by itself.
• It only changes queue order when decision_extraction_mode=hybrid_priority.

Decision Status Policy

Default:

• decision_default_status = draft
• decision_auto_confirm_enabled = false

Optional auto-confirm:

• Enabled only if decision_auto_confirm_enabled = true
• Requires confidence >= decision_auto_confirm_min_confidence

LLM Output Contract

The classifier expects strict JSON:

{
  "label": "decision | not_decision",
  "confidence": 0.0,
  "summary": "1-2 lines",
  "reason": ["bullet 1", "bullet 2"],
  "tags": ["optional-tag"]
}

Operational Notes

• If no LLM provider key/config is available, decision extraction is skipped safely and retried later.
• Activity logging is predictable and low-cost.
• Decision extraction cost is controlled by:
  • decision_batch_size
  • decision_backfill_days
  • decision_extraction_mode (llm_only vs hybrid_priority)