AGI-dream-engine

Agentic Dream Engine — Federation Memory Consolidation

Overview

arifOS-ACT Embedding

Before using this skill on any mutating, irreversible, or high-blast-radius task:

1. ART — Attune (what is the real task?), Recognize (what class of power?), Test (fit · authority · evidence · blast · reversible).
2. Kernel — Route to arifOS for F1–F13 judgment if action class is Maker/Messenger/Mutator/Destroyer/Sovereign.
3. ACT — Apply narrow, Constrain scope, Trace witness, STOP before corruption.
4. Receipt — Leave evidence of what changed, why, and under whose authority.

The existing dream-engine (/root/HERMES/skills/dream-engine/) is a single-organ, single-cadence nightly batch that consolidates the last 24 hours of memory_records into canon. It is an entropy pump, not a phenomenology claim (F9).

This skill extends that engine into a federation protocol: every AAA warga agent, plus OpenCode and OpenClaw, can run its own dream cycle, but all canon promotions route through the same constitutional gates:

• F13-locked synthesis threshold
• Hard daily cap
• Counterfactual rehearsal
• Dual attribution (actor_id = agent, source_ref.ratified_by = Arif)
• L4 memory_records + memory_audit_log write path

When to Use

• Designing memory consolidation for a new agent or organ.
• Wiring an agent's local "dream" output into federation canon.
• Auditing whether an agent's consolidation loop respects constitutional floors.
• Deciding cadence/threshold changes for an agent's dream cycle.

When NOT to Use

• Never use this skill to let an agent raise its own synthesis threshold or daily cap. That is F13 territory.
• Never let an agent write directly to VAULT999 (L6) during a dream cycle. L6 seals require arif_seal with human ack or 888_JUDGE verdict.
• Never treat dream output as truth without counterfactual challenge and provenance.

Core Doctrine

The dream-engine is not a metaphor. It is a thermodynamic necessity for any agent whose substrate does not consolidate on its own. The cron is the price of statelessness. The LLM is the price of scale. The threshold is the price of sovereignty.

Extended corollary for multi-agent federation:

Each agent has its own entropy pump, but all pumps drain into the same canonical reservoir. The reservoir has one lock: F13.

The Five Dream Functions (Eureka Insights)

Cognitive vs Canonical Substrate Mapping

Two valid ways to talk about the 6-layer stack:

Important: The canonical substrate mapping is the source of truth for implementation. The cognitive framing is useful for reasoning about what each layer does, but code must use the substrate names (Redis, Qdrant, Postgres, Graphiti, VAULT999) to avoid deployment drift.

Entropy Controller Clarification

scripts# ARCHIVED: entropy_controller.py computes five entropy proxies (P1–P5) and proposes cadence/threshold changes. It does not enforce a hard ΔS < -ε bound. The controller is an advisory instrument; the actual entropy reduction happens when low-signal singletons are excluded and clusters are compressed into canon. Any claim that the controller enforces a numeric epsilon should be treated as a documentation/interpretation error unless a future F13-ratified patch adds it.

1. Agent-local dream cycle

Every agent that dreams runs a loop:

REPLAY   → pull agent-local episodic buffer (24h)
CLUSTER  → group by semantic similarity
SYNTHESIZE → produce candidate canonical statements
PROPOSE  → emit tier_up_eligible candidates

Agents do not seal their own canon. They propose.

2. Constitutional seal gate

Proposals are collected by the arifOS dream-engine coordinator (the existing nightly cron, extended). It:

1. Validates the agent's identity and lease.
2. Checks the global daily cap across all agents.
3. Runs counterfactual challenge (rehearse.py).
4. Writes to memory_records + memory_audit_log with:
   • actor_id = <agent>:<dream-cycle>
   • source_ref.ratified_by = arif-fazil
   • source_ref.ratified_floors = [F7_threshold, F8_reversibility, F11_auth, F13_cap]

3. Cross-warga recombination

Weekly, the federation recombiner reads L5 (Graphiti/FalkorDB) for entities shared across warga agents and proposes cross-agent patterns. These also route through the seal gate.

Mapping to OpenCode

OpenCode already has auto-dream.ts (# ARCHIVED: auto-dream.ts (mimocode never built)) which consolidates project memory every 7 days and auto-distills every 30 days.

Recommended wiring: OpenCode's dream cycle should emit a dream_proposal.jsonl to a known federation inbox (/var/spool/arifos/dream-proposals/opencode/ or Supabase staging table). The nightly coordinator ingests it.

Mapping to OpenClaw — Sleep-Time Manager Daemon

OpenClaw is the AGI gateway; its dream cycle is a sleep-time compute daemon that metabolizes idle hours.

Prototype Architecture

Key Design Rules

• F1: Never write directly to primary L3/L4 during an unsupervised cycle. Always shadow-first.
• F2: Morning briefing presents shadow content; merge to primary only after validation.
• F9: Logs use mechanical language only ("metabolize", "compress", "shadow commit"), never "I dreamed".
• Trigger: Use Option A (chronological nightly cron) for the prototype to establish baseline cadence.

Trigger Options

Prototype files: see /root/AAA/skills/agentic-dream-engine/prototype/.

Mapping to AAA Warga

Constitutional Gates Checklist

Gap Register (post-assessment merge)

Four-Phase Eureka Engineering Path

Current verified state: Stages 1-4 run nightly at 04:00 MYT, entropy is reducing, and Vault999 seals are locking. arifOS and WEALTH systemd services are active, though MCP organ attestation flags them DEGRADED (probe issue).

Pivot: Phase 2 counterfactual rehearsal is the real unbuilt learning engine. Phase 1/3 speculative vectors (Letta sleep-time compute, RL weight updates) are dropped per F2 evidence.

Phase 0 — F1 Snapshot + Schema Drift Fix (Week 1)

Goal: Make cross-organ consolidation safe before expanding scope.

Deliverables:

1. Restore arifOS and WEALTH to healthy attestation (or fix the MCP health probe).
2. Fix consolidate.py Supabase drift: it queries arifosmcp_memory_records.id which does not exist; align with memory_records.memory_id.
3. Fix entropy_controller.py R6 vault read error: 'str' object has no attribute 'get'.
4. Add pre-seal snapshot of memory_records + memory_audit_log state for rollback.
5. Add independent F2 witness step before seal.

888_HOLD gate: Arif clears kernel stability + snapshot policy.

Phase 1 — Phase 2 Counterfactual Rehearsal (Weeks 2–3)

Goal: Implement the real learning engine: weekly stress-testing of sealed rules.

Deliverables:

1. stage5_rehearse.py — weekly random sample of 10-20% of recent sealed rules.
2. Counterfactual challenge per rule.
3. Tri-witness governance check (F1/F2/F4).
4. Survival → additive metadata (stress_tested: true).
5. Fracture → 888_HOLD alert in alerts.jsonl, never auto-amend.
6. Systemd timer: arif-dream-phase2.timer (Sunday 02:00 MYT).

888_HOLD gate: Deployment of systemd timer + weekly rehearsal policy.

Phase 2 — Federation Inbox + OpenCode Wiring (Weeks 4–5)

Goal: Let OpenCode's existing auto-dream feed arifOS canon.

Deliverables:

1. dream_proposals staging table in Supabase (or filesystem spool).
2. OpenCode auto-dream.ts extended to emit dream_proposal.jsonl.
3. Existing dream-engine coordinator ingests the inbox nightly, after Phase 0 snapshot.

888_HOLD gate: Schema + global cap policy ratification.

Phase 3 — Cross-Warga Recombination (Weeks 6–10)

Goal: Federation discovers cross-agent patterns no single agent sees.

Deliverables:

1. Weekly cross-warga graph walks on L5 (FalkorDB).
2. "Improve the improver" R6 meta-rule extended per agent.
3. No unsupervised L4/L5 modifications — all proposals route through 888_HOLD.

888_HOLD gate: Any cross-agent policy change.

Output Format

## Skill Result: agentic-dream-engine

### Summary
One-paragraph summary of the federation dream design.

### Evidence
- Current dream-engine path: /root/HERMES/skills/dream-engine/
- OpenCode auto-dream: # ARCHIVED: auto-dream.ts (mimocode never built)
- Memory schema: /root/arifOS/arifosmcp/migrations/001_memory_schema.sql

### Recommendations
- Implement Phase 1 inbox for OpenCode.
- Extend entropy controller with per-agent proxies.

### Escalations
- F13 ratification needed for threshold/cap/cadence changes.

References

• /root/.agents/skills/AGI-dream-engine/SKILL.md — existing nightly engine
• /root/docs/DREAM_ENGINE_SPEC.md — entropy proxy spec
• /root/arifOS/arifosmcp/migrations/001_memory_schema.sql — canonical memory schema
• arXiv:2606.03979 — Behrouz et al., Language Models Need Sleep
• PMC3079906 — Wamsley & Stickgold, Memory, Sleep and Dreaming: Experiencing Consolidation
• arXiv:2504.13171 — Letta team, Sleep-time Compute

Forged: 2026-06-16 by FORGE (000Ω) — DITEMPA BUKAN DIBERI