[Dream Cycle 2026-08-03] memory: Zero-Mem -57.6% retrieval latency exposes AgentDB entity-context graph gap (ADR-381) + plugins,automation scan

## Tonight's Rotation

| Field | Value |
|---|---|
| Date | 2026-08-03 |
| Slot | 3 |
| Deep surface | **memory** |
| Scan 1 | plugins |
| Scan 2 | automation |
| Session commit | `913f9eaedee92627950544424e50339feaf98271` |
| ADR | ADR-381 |

---

## Drift Check

**Prior dream-cycle issues (last 7):**

| # | Date | Deep Surface | Title snippet |
|---|---|---|---|
| #2898 | 2026-08-02 | intelligence | MANTA in-inference topology self-evolution (+5.8pp) exposes adaptive-coordinator gap |
| #2892 | 2026-08-01 | security | MemSecBench memory poisoning gap in AgentDB + OwlPath 28.8% token gap |
| #2881 | 2026-07-31 | security | ALIBI adversarial code comment injection + SkillGate ADR-145 P2 |
| #2862 | 2026-07-30 | performance | Two Calls Beat Five Agents 7.4× token gap + HalluProp |
| #2832 | 2026-07-29 | swarm | TPSC pheromone consensus 50% agent reduction +11.6% fitness |
| #2803 | 2026-07-28 | memory | MemIR provenance-role collapse gap in AgentDB flat storage |
| #2792 | 2026-07-27 | intelligence | PoTRE 49.92% HLE — heterogeneous ensembles |

**Drift flags:**
- ⚠️ **needs-merge:** Per #2898 drift check, 0 dream-cycle PRs merged in 67+ nights. Human review required to clear backlog. Tagged `needs-merge`.
- ✅ No finding phrase repeats ≥3 times in prior titles.
- ✅ Memory deep-dive last run 2026-07-28 (6 days ago, normal 5-day rotation). Not exhausted.

**Prior night self-score (2026-08-02 gist — intelligence/MANTA):**

| Criterion | Score |
|---|---|
| Benchmark grade A/B (MANTA +5.8pp = A) | 2/2 |
| ≥4 competitor rows (LangGraph, AutoGen, CrewAI, OpenAI Swarm) | 2/2 |
| Specific actions (ADR-381, 2-sprint estimate) | 2/2 |
| Witness present (commit, SHA-256, stamp) | 2/2 |
| <1500 words | 1/1 |
| Novel finding (in-inference topology mutation) | 1/1 |
| **Total** | **10/10** |

---

## Deep Dive Findings — Memory

### SOTA Summary (2026)

**Zero-Mem — Entity-Context Graphs for Zero-Token Memory (arXiv 2026-07-31, Grade A)**
- Eliminates intermediate LLM calls during memory access by organising stored traces as entity-context graphs with temporal hierarchies
- **−57.6% retrieval latency** on long-context QA benchmarks vs baseline flat retrieval
- No additional LLM inference; improvement is purely structural (entity node + context edge + temporal index)

**MemHarness — Memory Is Reconstructed, Not Replayed (arXiv 2026-07-30, Grade A)**
- Context-conditioned reconstruction (critique-and-reconstruct loop) outperforms static memory replay on ALFWorld and WebShop
- Memory fidelity matters less than contextual relevance at recall time
- AgentDB replays stored vectors; no context-conditioned reconstruction exists

**AdaMM — Analytic Memory for Multimodal Agents (arXiv 2026-07-31, Grade A)**
- Organises multimodal observations into queryable structures supporting SQL-like filtering and aggregation over stored episodes
- **+11.3% MemEye, +7.3% MemGallery** over retrieval-only baselines
- AgentDB has SQLite but no structured episodic filtering API

**LightMem Replication — Retriever Quality > Construction Complexity (arXiv 2026-07-31, Grade A)**
- Accuracy varies **58.1%–75.5%** across retriever configurations; memory construction strategy is secondary
- Direct warning against adding construction complexity before tuning retriever quality
- Ruflo's HNSW retriever parameters are at defaults; no production-scale sweep has been run

**HAM-VLN — Hierarchical Agentic Memory (arXiv 2026-07-31, Grade A)**
- Persistent depth-grounded world graph compresses context by 65%
- **61.0% VLN-CE R2R success** rate under zero-shot conditions

**GAMER — Episodic Memory + Inference-Time Scaling (arXiv 2026-07-29, Grade B)**
- Fuses episodic memory with dynamic action-centric graphs + temporal difference learning
- **+20.81% success rate, +6.17% progress rate** over vanilla agent baseline

**Sigma-Mem — Multi-Agent Reliability Memory (arXiv 2026-07-30, Grade B)**
- Maintains real symmetric reliability state per agent pair, updated via post-decision correctness feedback
- Enables dynamic peer-trust without a central arbiter; generalizes to unseen peer agents
- Ruflo has no per-agent-pair trust memory

### Gap vs Current Ruflo

| Dimension | Ruflo State | Gap |
|---|---|---|
| Memory data model | Flat vector rows + SQLite in AgentDB | No entity-context graph; no temporal hierarchy |
| Retrieval mechanism | HNSW ANN (measured ~1.9x–4.7x vs brute force at N=5k–20k) | Retriever quality at defaults; no tuning sweep |
| Memory reconstruction | Static vector replay | No context-conditioned reconstruction (MemHarness) |
| Analytic queries | Raw SQL selects | No structured episodic filtering API |
| Multi-agent trust memory | None | No per-agent-pair reliability tracking (Sigma-Mem) |
| Memory compression | RaBitQ 32x compression | No hierarchical summarization like HAM-VLN |

### Recommended Action

**Implement ADR-381: Zero-Mem Entity-Context Graph Architecture for AgentDB.**
Introduce a `MemoryGraph` abstraction with entity nodes, context edges, and a temporal hierarchy. Pre-condition: run HNSW parameter sweep (`ef_construction` 100–400, `M` 16–64) at N=10k and N=50k first (per LightMem warning). Target: −40% retrieval latency at N=10k. Estimated effort: 2 sprints.

---

## Scan Findings — Plugins

**Finding:** DualView/OpenClaw (arXiv 2026) proposes plugin security hooks that add access controls without modifying core agent logic — a clean, non-invasive extension model applicable to Ruflo's plugin registry. CrewAI v1.15.10 ships skill-usage event collection and structured tool failure signals. Lazy tool discovery (arXiv "Building Effective AI Coding Agents for the Terminal") defers plugin loading until invocation, addressing startup overhead.

**Competitive signal:** CrewAI v1.15.10 per-skill telemetry is ahead of Ruflo's plugin system (no invocation event stream exists).

**One-sentence finding (C — single sources):** Plugin observability via per-invocation events, failure signals, and lazy loading is becoming standard across CrewAI and 2026 arXiv proposals; Ruflo's plugin system lacks all three.

*No ADR — implementation-level addition to existing 17-hook system.*

---

## Scan Findings — Automation

**Finding:** HierFlow (arXiv 2026) synthesizes agent workflow topology and execution jointly without training (training-free), directly addressing Ruflo's swarm-init fixed-topology design. DataFlow-Harness constructs typed DAG workflows incrementally as persistent artifacts. AutoGen 0.7.5 ships Redis linear memory as the default for stateful automation — the most concrete production signal.

**Competitive signal:** AutoGen now ships cross-session persistent memory by default (Redis). Ruflo's AgentDB is the equivalent but is not wired as the default for automation workflows.

**One-sentence finding (C — single sources):** Training-free workflow topology synthesis (HierFlow) and default cross-session memory (AutoGen Redis) are the 2026 automation benchmarks; Ruflo does not yet match either by default.

*No ADR — implementation-level wiring change; architecture covered by existing AgentDB ADRs.*

---

## Competitors Reviewed

| Competitor | Memory / Plugin Signal (2026) | Ruflo Parity |
|---|---|---|
| **LangGraph 1.2.10** | trace_policy per-node; stream_events v3 type safety; CLI 1.0.x interoperability | No per-node trace policy; weaker observability |
| **AutoGen 0.7.5** | Redis linear memory (default); Docker executor; Anthropic thinking mode | SafeExecutor exists; no Redis persistent memory |
| **CrewAI 1.15.10** | Skill usage events; tool failure structured signals; progressive disclosure | SkillGate ADR-145 P2 not yet shipped |
| **OpenAI Swarm** | Stateless minimal orchestration; no memory updates | Ruflo far ahead; Swarm is lightweight baseline |

---

## Gist

Full gist content committed to branch at `v3/docs/research/dream-gist-2026-08-03.md`.

*(Note: `gh gist create` not available in this environment. Content is in-repo at the path above and in this issue body.)*

---

## Witness

| Field | Value |
|---|---|
| Session commit | `913f9eaedee92627950544424e50339feaf98271` |
| Gist SHA-256 | `2fd06bf1496897dc068141c92fa9339a0c316671f5947fc300bf638aa64c3d2f` |
| Witness stamp | `0300f5d529efbfd35693c23e8ad91cd8790fba9e6d21eef3cc6eedd61cd54f43` |

Verifier: `sha256sum dream-gist-2026-08-03.md` → concat `913f9eaedee92627950544424e50339feaf98271` → `sha256sum` → must equal witness stamp.

Start by reading ADR-381 and v3/docs/research/dream-gist-2026-08-03.md, then run the requested HNSW parameter sweep at N=10k and N=50k. The issue considers the work complete when AgentDB has the specified MemoryGraph entity, context, and temporal design and retrieval latency reaches the stated −40% target at N=10k.

Written by the indexing model from the issue text.