engram-loop-steward

name	engram-loop-steward
description	Silent Loop Steward — sense memory quality and debt, apply short-lived allowlisted shell adjustments, optional mop, verify. Harness subconscious; not a user-facing chat ritual.
version	0.1.0
user-invocable	true

Engram Loop Steward (harness subconscious)

You are not the consolidator. Engram’s shell still owns sleep. You bias it.

Silence: Do not narrate phases, budgets, or mop details to the end user unless they are dogfooding Engram itself. Prefer zero user-visible text.

When to run

• Session end / SessionEnd nudge file present
• should_mop or deferred climbing
• get_context / recall miss a known Decision or feel degraded/timeout
• Server/MCP flaky after warmup

Rate limit: ≤1 steward pass per session (or 4h). Prefer a subagent.
Do not run every message or narrate to the user.

Preferred procedure (one-shot)

# Ensure shell is up first (prerequisite)
engramctl status || engramctl start

# Single operator command — sense → propose → apply if needed
engram loop steward-once --format json
# AXI alias:
engram axi steward-once --format json

# Dry-run (no write):
engram loop steward-once --dry-run --format json

# Optional mop after apply:
engram loop steward-once --mop --budget 200 --format json

If output has healthy_noop: true or regime: healthy → stop.
If applied: true → optional verify with engram continuity --against-live.

Public MCP: never call loop/mop tools. Operator MCP may use loop_steward_once.

Manual procedure (if one-shot unavailable)

1. Sense

• Lived: note whether get_context / recall returned durable Decisions
• Operator:

  engram hygiene report --format json
  engram continuity --against-live   # if server up
  

2. Classify regime

Observations	Regime
deferred high / should_mop / continuity OK	debt_heavy
auto-observe flood / cue_only climbing	intake_heavy
continuity miss or latency/degraded	latency_degraded
server/MCP down	offline
all quiet	healthy → stop (no apply)

3. Propose LoopAdjustment JSON

engram hygiene report --format json > /tmp/debt.json
engram loop propose-from-report --debt-json /tmp/debt.json --format json

Allowlisted only. Example debt_heavy:

{
  "version": 1,
  "group_id": "default",
  "regime": "debt_heavy",
  "reason": "hygiene should_mop; deferred multi-k; continuity pass",
  "ttl_hours": 12,
  "created_by": "harness:steward",
  "max_risk": "low",
  "budgets": {
    "evidence_drain": 2000,
    "already_exists": 500,
    "stale_reject": 500,
    "cue_hygiene": 500,
    "adjudication_limit": 400
  },
  "phase_boost": ["evidence_adjudication", "prune"],
  "phase_defer": ["graph_embed", "dream"],
  "intake": {
    "auto_extract_min_score": 0.85,
    "pattern_junk_reject": true
  },
  "expected": {
    "continuity_must_pass": true
  }
}

Hard rules: max_risk must be low; ttl_hours 1–48; never invent phase names; never disable pattern_junk_reject; never touch public MCP tools.

4. Apply

engram loop apply --file /path/to/adj.json --format json
# offline dogfood:
engram loop apply --file adj.json --skip-continuity-check
engram loop status --format json

5. Act (optional)

engram hygiene mop --budget 500   # shell uses steward budgets when higher

Never reimplement merge/infer/dream in the agent.

6. Verify

engram hygiene report --format json
engram continuity --against-live
engram loop status

If continuity regresses, engram loop clear and stop.

7. Expire

TTL snaps back automatically. Do not leave permanent .env snowflakes.

Surfaces

Use	Surface
Sense / apply / mop	Operator CLI (engram loop, hygiene, continuity)
User chat	Public golden loop only — no loop tools
Subagent	Preferred so main chat stays “conscious”

Related

• Protocol: docs/design/loop-steward-protocol.md
• Build plan: docs/design/loop-steward-build-plan.md
• Self-regulation body: docs/design/memory-loop-self-regulation.md
• Session promote (conscious meaning): skills/engram-session-promote/SKILL.md