[Dream Cycle 2026-05-27] intelligence: SR²AM 8B=120–355B via simulative planning — 95% token gap + capabilities,memory scan

## Tonight's Rotation

| Field | Value |
|-------|-------|
| **SLOT** | 2 |
| **DEEP surface** | intelligence |
| **SCAN surfaces** | capabilities, memory |
| **Session commit** | `733ada1b9aecd878418c88f0ae3e900523d59c37` |
| **Date** | 2026-05-27 |

---

## Drift Check

- **Prior dream-cycle issues found:** #2149 (2026-05-26, DEEP=security) — different surface, no phrase repetition
- **Prior self-score (issue #2149):** 9/10 — benchmark grade A/B ✓, ≥4 competitor rows ✓, specific actions ✓, witness present ✓, <1500 words ✓, novel finding ✓
- **No dream-cycle PR merged:** First run was 2026-05-26 (1 night ago — under 14-night threshold, no `needs-merge` flag yet)
- **Drift flags:** None — surfaces rotating normally

---

## Deep Dive Findings — Intelligence

### SOTA Summary (2026)

Three high-signal 2026 papers define the new intelligence frontier:

**SR²AM** (arXiv:2605.22138, Grade A): Self-Regulated Simulative planning decomposes agent decision-making into simulative reasoning (forward prediction), self-regulation (when to plan), and reactive execution. Key benchmark result: an 8B-parameter model matches systems with 120–355B parameters; a 30B model competes with 685B–1T systems at **25.8–95.3% fewer reasoning tokens**. RL training increased average planning depth by 22.8% while planning *frequency* grew only 2.0% — meaning the model plans selectively, not uniformly.

**MUSE-Autoskill** (arXiv:2605.27366, Grade A): Introduces a skill lifecycle (create → store → evaluate → refine) that treats skills as long-lived, experience-aware, testable assets. Evaluated on SkillsBench. Key value: cross-agent skill transfer — skills proven in one agent are reusable by others without re-learning.

**Agent Lifespan Engineering** (arXiv:2605.26302, Grade A): Identifies 4 degradation mechanisms for long-lived agents: compression aging, interference aging, revision aging, maintenance aging. Each requires a different diagnostic and repair strategy. SONA-class systems face all four if not explicitly managed.

### Gap vs Current Ruflo

Ruflo's 4-step intelligence pipeline (RETRIEVE→JUDGE→DISTILL→CONSOLIDATE) and 3-tier routing (ADR-026) have no simulative planning primitive. Every Tier-3 call commits to tool execution without a forward pass. This wastes tokens on simple subtasks and underinvests in planning for multi-step tasks. Additionally, SONA pattern matching is not a skill lifecycle — skills cannot be quarantined, promoted, or transferred cross-agent.

### Recommended Action

**ADR-131** (filed): Add `SimulativePlanningRouter` as conditional pre-execution layer in the `route` hook. Activated when estimated horizon > 5 steps or ≥2 MCP calls predicted. Uses Haiku as a world model for a ~256-token shadow pass (~$0.0002) before Tier-3 dispatch. Target: ≤30ms overhead, ≥20% token reduction on multi-step tasks.

---

## Scan Findings — Capabilities

**Source:** benchmarkingagents.com, arXiv:2503.16416v2 (survey), MarkTechPost (Apr 2026)

**Competitive signal (Grade B):** SOTA on major agent benchmarks 2026:
- SWE-bench Verified: Claude Opus 4.7 @ 87.6%, GPT-5.3 Codex @ 85.0%
- GAIA (Princeton HAL): Claude Sonnet 4.5 @ 74.6%
- WebArena: Claude Mythos Preview @ 68.7%

Note: top scores inflated 5–15 points by contamination/scaffolding — treat as directional.

**Finding:** Ruflo has no benchmark harness for agent capability regression detection. Adding `npx claude-flow performance benchmark --suite agent` (10-question GAIA subset) would enable nightly regression detection with zero architectural change. Implementation-level — no ADR needed.

---

## Scan Findings — Memory

**Source:** mem0.ai/blog/state-of-ai-agent-memory-2026 (Grade A)

**Competitive signal (Grade A):** Mem0 publishes 2026 production benchmarks:
- LoCoMo: 92.5 (multi-session recall)
- LongMemEval: 94.4 (knowledge updates)
- BEAM@1M tokens: 64.1 → BEAM@10M tokens: 48.6 (~25% degradation at scale)

Gains: +29.6 points on temporal queries, +23.1 on multi-hop since prior algorithm.

**Finding:** Ruflo's HNSW+AgentDB has no public benchmark score and no temporal abstraction layer. At 1M+ token scale, unaddressed temporal reasoning degrades ~25%. Production gap: no implementation of cross-session identity resolution or memory staleness eviction. These are implementation-level improvements to `@claude-flow/memory` — no ADR needed, but worth tracking.

---

## Competitors Reviewed

| Framework | Adaptive Depth Planning | Skill Lifecycle | SWE-bench Score | Memory SOTA |
|-----------|------------------------|----------------|-----------------|------------|
| **Ruflo (claude-flow)** | No (flat tier routing) | No (static types) | Not measured | HNSW+AgentDB (no public score) |
| **OpenAI Agents SDK** | Yes (o3 chain-of-thought orchestration) | Partial (handoff templates) | ~85% (GPT-5.3 Codex) | Platform-managed |
| **LangGraph v0.4** | No (static graph edges) | No | Not published | Checkpoint-based, no vector native |
| **CrewAI Enterprise** | No (sequential/hierarchical fixed) | Partial (role templates) | Not published | Via mem0 integration |
| **AutoGen 1.0 GA** | Partial (GroupChat dynamic routing) | No | Not published | Basic session memory |

---

## Gist Link

`v3/docs/research/dream-cycle-2026-05-27-intelligence.md` (committed to branch `dream/2026-05-27-intelligence`)

_No public gist URL — running without gh CLI gist support in this environment. Gist content committed to branch for auditability._

---

## ADR

**ADR-131** — `v3/docs/adr/ADR-131-simulative-planning-router.md` — Simulative Planning Router: Selective Depth Allocation for Agent Intelligence (Status: Proposed)

---

## Witness

| Field | Value |
|-------|-------|
| **Session commit** | `733ada1b9aecd878418c88f0ae3e900523d59c37` |
| **Gist SHA-256** | `20171013681d33297870fb922e666fc4bdef7ac07f65224f1a7f918e36a4c531` |
| **Witness stamp** | `96fc07c70df4d4fac0400031aa765583cb28e8940c80276602b4c124e65a4ed3` |
| **Verifier** | `sha256sum dream-cycle-2026-05-27-intelligence.md` → concat session commit → `sha256sum` → must equal witness stamp |

Read v3/docs/research/dream-cycle-2026-05-27-intelligence.md and v3/docs/adr/ADR-131-simulative-planning-router.md first. Then inspect the route hook and @claude-flow/memory references mentioned in the findings, along with the 53-comment discussion. Done requires a settled scope and acceptance criteria for the proposed router, benchmark, or memory work.

Written by the indexing model from the issue text.