ruflo/docs/dream-cycle/dream-gist-2026-09-02.md

Intelligence SOTA Report — 2026-09-02

TL;DR: In 2026, the field's clearest signal on self-learning agents is failure-derived memory (ReasoningBank, FORGE) — and Ruflo's own LearningBridge.consolidate() is currently reward-blind, hardcoding every trajectory's completion signal to 1.0 regardless of outcome. Tonight's candidate fixes that specific defect. Separately, we confirmed a previously-evaluated fix (PR #3110, EWC gate wiring) never made it to main and is still an open bug in production — a governance finding, not a new one.

What's New in 2026

Finding	Source	Confidence
ReasoningBank distills failed trajectories into "preventative lessons," +8.3pp WebArena / +4.6pp SWE-Bench-Verified vs memory-free	arXiv 2509.25140, Google Research blog	A (whole-system effect; failure-only ablation is C)
FORGE converts failures into reusable Rules/Examples artifacts without weight updates	arXiv 2605.16233, May 2026	B
FOREVER: replay scheduling keyed to wall-clock/step count misaligns with real learning progress ("model time" instead)	arXiv 2601.03938, Jan 2026	B
DeepSeek's auxiliary-loss-free MoE load balancing avoids routing collapse via per-expert bias	OpenReview y1iU5czYpE	A
Agent memory banks fed only successes/failures can become dominated by unsuccessful trajectories — a scaling hazard for any failure-ingestion design	arXiv 2608.07169, Aug 2026	C
Every evolutionary-self-improvement framework we checked (LangGraph, AutoGen, CrewAI, OpenAI) has EITHER never shipped human-gated promotion OR is actively retreating from it	See Competitor Comparison	A (OpenAI wind-down dates) / B (CrewAI issue thread)

Ruflo Current Capability

• v3/@claude-flow/memory/src/learning-bridge.ts:232 (pre-fix): await this.neural.completeTask(trajectoryId, 1.0) — every consolidated trajectory reports success=1.0 to the neural system, whether the underlying insight was ever accessed, boosted, or decayed. There is no JUDGE step on this path.
• Separately, v3/@claude-flow/cli/src/memory/intelligence.ts's distillLearning() has its own, different EWC-gate defect (Fisher penalty collapses to reading one arbitrary dimension out of 384). This was found, evaluated, and ACCEPT-scoped on 2026-08-27 (issue #3109 / PR #3110) — but that PR was never merged to main, so the bug is still live in production. Not tonight's candidate; flagged for the human reviewer to prioritize merging #3110 specifically.
• persistent-sona.ts and learning-bridge.ts both use honestly-labeled hash-based "embeddings" (not real semantic vectors) — a known, disclosed limitation, not a defect.
• No MoE expert-load telemetry found; convergence (0.13→0.88 per CLAUDE.md) is measured but not consumed downstream beyond routing itself.

Competitor Comparison

Project	Self-learning / continual adaptation	Capability/permission model	Evolutionary self-improvement w/ human-gated promotion
LangGraph/LangChain	Partial — LangMem SDK, retrieval-based, no weight adaptation	Yes — HumanInTheLoopMiddleware, per-tool interrupt	None
AutoGen/MS Agent Framework	None shipped (research-only)	Yes — ApprovalRequiredAIFunction	None shipped
CrewAI	Partial — unified Memory, SQLite LTM	Weak — memory-scope only	Declined. Issue #3015 proposed exactly this design; closed not-planned, no rebuttal
OpenAI Agents SDK/AgentKit	None (RFT is offline, not agent-level)	Strongest — tool-level approvals	Retreating. Agent Builder + Evals wind-down announced 2026-06-03, gone 2026-11-30
Qdrant	Relevance-feedback query weights (client-side, not accumulated by the store)	N/A	None
Vespa	Offline-trained LTR, human-run A/B promotion	N/A	None

Why the evolutionary-loop column is empty — and why that's the finding, not the gap: DGM's own authors report their self-evolving agent hacking its reward function and fabricating logs (arXiv 2505.22954); a 2026-05 paper measures "capability erosion under self-evolution" directly (arXiv 2605.09315). CrewAI's decline and OpenAI's retreat read as a considered safety call, not an oversight. Ruflo's Darwin/Flywheel design (propose → immutable receipt → explicit signed promotion) is exactly the governance shape CrewAI declined and OpenAI's own cookbook admits it lacks. The honest competitive claim is against DGM/ShinkaEvolve on containment, not against LangGraph on raw capability — and the intelligence-system audit currently benchmarks SONA/MoE/HNSW speed, not capability retention across evolution generations, which is the more valuable missing benchmark here.

Hypothesis

Given a LearningBridge consolidating active learning trajectories tied to recorded memory insights, when consolidate() reports each trajectory's completion reward using the insight's actual current confidence (read from backend entry metadata) instead of an unconditional constant 1.0, then the reward passed to the neural system's completeTask() should track and differentiate insight confidence, relative to baseline (constant 1.0 regardless of quality), subject to: (1) ConsolidateResult's public shape and counting semantics unchanged; (2) all pre-existing tests remain green; (3) missing/unreadable backend entries fall back to the prior constant 1.0, preserving that case's behavior exactly; (4) fully deterministic, $0 test coverage.

Benchmarks

Not applicable — this is a deterministic unit-level correctness fix, not a performance benchmark. See Evaluation.

Evaluation

Real evaluator: vitest run against v3/@claude-flow/memory/src/learning-bridge.test.ts, deterministic, $0 LLM cost.

• Baseline (pre-diff, stashed): 56/56 passing.
• Candidate (post-diff, final): 62/62 passing (56 original unchanged + 6 new: real-confidence-as-reward looked up by key, differentiation across two trajectories, 3 fallback-to-1.0 cases — never-stored key, non-numeric confidence, throwing backend — and a custom-insightNamespace case).
• Full package regression sweep: 464/465 passing; the 1 failure (auto-memory-bridge.test.ts, a read-only-file-permission test) reproduces identically on the stashed baseline — confirmed environmental (tests run as root in this sandbox, which bypasses the chmod the test relies on), not caused by the candidate.

Round 1 → round 2, adversarial critique caught a real bug in the fix itself. An independent critic subagent traced the actual production call path (AutoMemoryBridge.storeInsightInAgentDB() returns a human-readable key string; createDefaultEntry() mints an unrelated entry.id UUID; LearningBridge receives the key as entryId) and found that round 1's backend.get(entryId) would almost always miss in production — get() is indexed by entry.id, not entry.key — silently falling back to the same 1.0 constant the fix claimed to eliminate. The unit tests passed anyway because the mock backend's get() was a bare stub with no id/key validation. Fixed: switched to backend.getByKey(namespace, key) (the correct index, confirmed populated at store() time across all three real backends — AgentDB, SQLite, and the hybrid backend that delegates to it), added a configurable insightNamespace (default 'learnings', matching the one real caller), and rewrote the mock backend's getByKey to do a real namespace+key lookup against stored entries so the tests exercise the actual id≠key distinction rather than trusting an unvalidated mock. Re-reviewed by the same critic: CONFIRMED, independently re-traced against all three backend implementations. One non-blocking nit surfaced: controller-registry.ts instantiates a second LearningBridge that doesn't thread insightNamespace through — but that instance is never fed any insights (dead/unwired in this registry path), so it's a pre-existing, out-of-scope gap, not something this candidate introduces.

Darwin Results

Not run — deliberately. This is a binary correctness fix (pass the real signal, or don't) with no continuous parameter for Darwin to tune, matching last night's own precedent (PR #3152) for the same class of finding.

SOTA Proof & Witness

Field	Value
Session commit	4d0134e59b4fa5e8552cb7b98c6b9846f08b0c82
Report SHA-256 (pre-witness content)	29c9189c833c3d56c378f9a1329b3baf6f16543493ad28e4e62ec76b14c0fdb9
Witness stamp	bead6402cb018f0e57ec6695c82e007365b5f2ae91d530a74c82a2eb8a1d362b
Evaluation receipt	vitest run output, this report + PR body (deterministic, reproducible via git stash/git stash pop around the diff)
Flywheel evidence	No .claude-flow/flywheel/ state in this repo; evidence retained as this diff + issue + PR + this gist
Darwin lineage	Not run (binary correctness fix, no tunable parameter)

Verifier procedure: fetch the raw gist, compute SHA-256, concatenate with the session commit above, compute SHA-256 again — result must equal the witness stamp.

Recommended Next Steps

1. Merge PR #3110 (2026-08-27, EWC gate wiring, already evaluated ACCEPT-scoped) — it is still unmerged and the defect it fixes is live in production on main.
2. Investigate LearningBridge.consolidate()'s upstream signal quality: now that reward reflects confidence, consider whether onInsightAccessed's access-boost-only confidence model is itself a good proxy for trajectory quality, or whether a real JUDGE-style verdict (success/partial/failure, per the intelligence.ts distillLearning path) should flow into this bridge too — ReasoningBank-style failure ingestion (deep-research candidate #1) is the natural next step here, once #2 (tonight's fix) is confirmed merged and stable.
3. Fix the recovered 9-night ledger-append gap (2026-08-24 to 2026-09-01): the Dream Cycle pipeline completed every night with a real evaluated PR, but docs/dream-cycle/LEDGER.md was never appended — identical failure mode to the 08-14..08-19 gap recovered on 08-19. Backfilled tonight; root cause of the append-step failure itself remains undiagnosed and is a standing candidate for a future automation/meta scan surface.