engram-session-promote

name	engram-session-promote
description	End-of-session sparse promotion — propose at most 5 high-signal Decisions/Preferences/People for Engram remember(), never dump recaps.
version	0.1.0
user-invocable	true

Engram Session Promote (sparse)

Run this at session end, before context compaction, or when the user says "promote what matters" / "save durable decisions."

Product metric is multi-agent continuity — not exhaustively storing the chat.

Hard rules

1. Cap: 0–5 remember() calls per compaction window (not multi-day session).
2. Prefer 0 when nothing is durable. Silence is fine.
3. Never promote session recaps ("what we did today", transcript dumps).
4. Types only: Decision, Preference, Person, Correction, Goal, Commitment.
5. You are the extractor: always pass proposed_entities + proposed_relationships with a verbatim source_span that appears in content.
6. Pass model_tier (opus/sonnet/haiku). After PreCompact, pass compaction_id if the harness exposed one (or rely on ~/.engram/promotion-window.json).

Procedure

1. Scan this conversation for portable facts that another agent would need cold (without a handoff doc).
2. Rank by durability:
   • Identity / explicit preferences / corrections
   • Product decisions with lasting consequences
   • Named people the user cares about
   • Active goals / commitments with owners
3. Keep the top ≤5. Drop noise, temporary task status, and file paths that are only true inside this repo checkout.
4. For each kept fact, call one remember() with short content that contains the span:

remember(
  content="<1–3 sentences including the exact decision text>",
  model_tier="sonnet",
  proposed_entities=[{
    "name": "<atomic decision or person name>",
    "entity_type": "Decision|Preference|Person|Correction|Goal|Commitment",
    "source_span": "<verbatim substring of content>",
    "summary": "<one-line why it matters later>"
  }],
  proposed_relationships=[{
    "subject": "<entity or project>",
    "predicate": "DECIDED|PREFERS|RELATED_TO|CORRECTS|COMMITS_TO",
    "object": "<other entity or decision name>",
    "source_span": "<verbatim substring of content>"
  }]
)

5. Read the response:
   • committed_entities / committed_relationships → promotion landed
   • identity_core: true → protected for merge/prune
   • status: rejected + promotion_window_cap → stop; remaining facts wait for the next compaction window
6. Optionally call get_context once and confirm ≥1 promoted Decision surfaces.
7. Tell the user briefly: N promoted / M skipped (no essay).

What not to promote

• Long chat summaries or "session notes"
• Transient bugs already fixed in-repo
• Raw log lines, stack traces, or secrets
• Cadence / work-ticket noise unless the user explicitly wants it portable
• More than five items "just in case"

When uncertain

Use observe() for bulk/uncertain context, or skip.
If uncertain whether it is durable — do not remember it.

Related

• Golden loop: docs/GOLDEN_LOOP.md
• Main memory skill: skills/engram-memory/SKILL.md
• PreCompact window reset: hooks/pre-compact.sh