# AGI-dream-engine

> Extend the arifOS dream-engine so every AAA warga (333-AGI, 555-ASI, 888-APEX, A-AUDIT, A-ARCHIVE), OpenCode, and OpenClaw can autonomously consolidate memory without violating F1-F13.

- Skill: `ariffazil/agi-dream-engine` (Agent Skill, multi-file: 14 files)
- Install (CLI): `npx skillmds@latest add ariffazil/agi-dream-engine`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ariffazil/agi-dream-engine/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: ariffazil (https://skillmd.com/u/ariffazil)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/ariffazil/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)

| Function | Biology analog | Engineering analog | Current owner | Federation extension |
|----------|---------------|-------------------|---------------|----------------------|
| **Replay** | Hippocampal reactivation | Pull recent episodic memory | dream-engine stage1 | Each agent replays its own trajectory store |
| **Cluster** | Sharp-wave ripple binding | Cosine grouping | dream-engine stage2 | Agent-local + federation cross-agent cluster |
| **Synthesize** | Gist extraction | LLM compression to falsifiable statement | dream-engine stage3 | Per-agent + cross-warga synthesis |
| **Defend + Seal** | Counterfactual threat rehearsal | Counterfactual challenge before canon write | dream-engine stage4 + rehearse.py | Centralized via arifOS kernel |
| **Recombine** | Creative interleaving (Kekulé/Mendeleev) | Cross-organ graph walk | dream-engine recombine.py | Weekly cross-warga pattern detection |

### Cognitive vs Canonical Substrate Mapping

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

| Layer | Cognitive framing (intelligence ladder) | Canonical substrate framing (arifOS memory skill) |
|-------|----------------------------------------|---------------------------------------------------|
| L1 | Working context | Redis ephemeral |
| L2 | Short-term/episodic | Redis session / raw logs |
| L3 | Semantic/facts | Qdrant vector space |
| L4 | Procedural/skills | Postgres `memory_records` + `memory_store` |
| L5 | Constitutional/values | Graphiti/FalkorDB relational graph |
| L6 | Core identity/legacy | VAULT999 immutable ledger |

**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.

| OpenCode concept | Federation mapping | Gap |
|------------------|-------------------|-----|
| `AUTO_DREAM_TITLE` session | Local replay + cluster + synthesis | Not yet writing to arifOS L4 |
| `AUTO_DISTILL_TITLE` session | Pattern packaging into skills | Should produce canon candidates, not just skills |
| `Config.dream.interval_days` | Agent-local cadence | Must also respect federation global cadence/cap |

**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

| Stage | Function | Reversibility Mechanism |
|-------|----------|------------------------|
| **1. State Monitor** | Detect idle (`t_idle > 120 min` or scheduled nightly window); acquire L1 mutex | Suspends immediately on 888 activity |
| **2. L2 Ingress & Entropy Sorting** | Pull raw L2 traces; cluster semantically; drop noise | Read-only on primary stores |
| **3. Offline EMD Metabolizer** | Gist extraction + anticipatory reasoning trees | Output stays in shadow partition |
| **4. Shadow Commit & Seal** | Write compressed gist to `L3_shadow`; cryptographic hash; morning briefing handshake | Single-command purge or merge |

### 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

| Option | Mechanism | Pros | Cons | Recommendation |
|--------|-----------|------|------|----------------|
| **A — Chronological** | Fixed nightly window (e.g., 03:00 MYT) | Predictable cost; guaranteed morning briefing | Wasted compute on light days; may clash with late sessions | **Prototype** |
| **B — Saturation** | Trigger when L2 token/complexity threshold crossed | Compute-efficient; dreams only when material exists | Unpredictable timing; may miss briefings | Phase 2 |

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

## Mapping to AAA Warga

| Warga | Dream content | Output |
|-------|--------------|--------|
| **333-AGI** | Code forge receipts, plan DAGs, test results | Engineering canon |
| **555-ASI** | Memory synthesis, ethical critiques, cross-agent patterns | Memory + dignity canon |
| **888-APEX** | Verdict patterns, floor violations, seal chains | Governance canon |
| **A-AUDIT** | Anomaly patterns, drift signals | Audit canon |
| **A-ARCHIVE** | Vault chain topology, seal topology | Archive canon |

## Constitutional Gates Checklist

| Floor | Requirement | Current Status |
|-------|-------------|----------------|
| F1 AMANAH | Proposals only; no agent self-seals. Rollback = `DELETE FROM memory_records WHERE actor_id = '<agent>:<cycle>'`. **Add pre-seal snapshot of `memory_records` + `memory_audit_log` state so hallucinated L3/L4 writes can be reverted cleanly.** | Partial — additive writes exist, but snapshot rollback is missing |
| F2 TRUTH | Every canonical statement is falsifiable and carries a counterfactual. **Add explicit tri-witness cross-check before seal (human/AI/earth or second-agent verifier).** | Partial — counterfactual is generated, but no independent witness validates offline replay |
| F4 CLARITY | One candidate = one cluster = one proposition. No noise promotion. | Aligned |
| F7 HUMILITY | `synthesis_score ≥ 0.65` is F13-locked; no agent overrides. | Aligned |
| F8 LAW | Agent boundaries enforced via arifOS lease + A2A auth. | Aligned for Hermes; gaps for OpenCode/OpenClaw/warga |
| F9 ANTIHANTU | Logs are process descriptions, never "I dreamed". | Aligned |
| F11 AUTH | Every write has `actor_id`, `session_id`, and `memory_audit_log` row. | Aligned |
| F13 SOVEREIGN | Threshold, daily cap, and cadence changes require Arif's explicit ratification. | Aligned |

### Gap Register (post-assessment merge)

| ID | Gap | Severity | Owner | 888_HOLD gate |
|----|-----|----------|-------|---------------|
| G1 | No pre-seal L3/L4 snapshot for clean rollback | High | dream-engine | Before Phase 0 execution |
| G2 | Offline replay lacks independent F2 witness before seal | Medium | dream-engine stage4 | Before Phase 1 federation inbox |
| G3 | OpenCode auto-dream does not write to federation canon | High | OpenCode / AAA | Phase 1 ratification |
| G4 | OpenClaw and AAA warga have no dream cycle | High | AAA / OpenClaw | Phase 2 design |
| G5 | `entropy_controller.py` computes proxies but does **not** enforce a hard ΔS < -ε | Clarification | dream-engine | Document, do not change without F13 |
| G6 | arifOS + WEALTH degraded; global recombine unsafe | Blocker | ops | Phase 0 repair |

## 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*