docs/dream-cycle/2026-07-08-memory-sota.md
TL;DR: 2026 papers unanimously show passive HNSW retrieval is giving way to RL-navigated multi-granularity memory pyramids; Ruflo AgentDB remains a passive store and needs an active navigation layer to match the new SOTA.
| Finding | Source | Confidence |
|---|---|---|
| NapMem: RL-navigated multi-granularity memory pyramid beats passive retrieval on long-horizon tasks | Xu et al., arXiv 2026-07-06 | B |
| StateFuse: CRDT conflict-preserving memory enables auditable multi-agent state divergence | Volkov et al., arXiv 2026-07-07 | B |
| Memory in the Loop: in-process microsecond stores eliminate redundant agent actions vs disk-backed retrieval | Khan & Lipizzi, arXiv 2026-07-06 | B |
| MRMS: 3-temporal-axis (short/medium/long-term) unified memory substrate for long-lived agents | Li & Shi-Nash, arXiv 2026-07-05 | B |
| Mem0 2026: fused scoring (semantic + BM25 + entity) yields +29.6 pts temporal, +23.1 pts multi-hop vs prior | mem0.ai, 2026 | B (vendor) |
| Sovereign Memory Stack: L1 Redis <1 ms, L2 Qdrant HNSW+BQ 20 ms p99, L3 Pinecone episodic | Ranksquire, verified Mar 2026 | B |
| Princeton NLP: single agent matches multi-agent on 64% of benchmarked tasks at ~half the cost | presenc.ai/research, 2026 | C (secondary) |
| 12–20% of agent marketplace skills found malicious in public registry audits | Tony Kipkemboi, 2026 | C (single source) |
| Component | Current State | Gap |
|---|---|---|
| AgentDB storage | sql.js SQLite + HNSW (measured 1.9×–4.7× vs brute force at N=5k–20k) | No active RL navigation |
| Retrieval | Passive HNSW vector similarity (384-dim ONNX) | No fused BM25+entity+semantic scoring |
| Multi-agent state | Each agent writes independently; no conflict object model | No CRDT divergence surface |
| Memory tiers | Single layer (AgentDB); no explicit short/medium/long-term axis | No MRMS-style temporal tiering |
| Plugin security | IPFS registry with trust levels; no runtime malice scanning | No behavioral audit at install time |
| Working memory | Context window + retrieval; no in-process sub-millisecond store | Potential latency overhead on high-frequency queries |
| Framework | Memory Architecture | 2026 Notable Update | GitHub Stars |
|---|---|---|---|
| LangGraph | State graph + reducer logic + DeltaChannel | Per-node timeouts, typed streaming v2 | ~35K |
| CrewAI | Pluggable RAG/knowledge/vector backends | v1.14: Snowflake Cortex, pluggable memory (May 2026) | ~29K |
| MS Agent Framework (AutoGen + SK) | Stateful + MCP + A2A native | Merged Apr 2026; unified .NET + Python | ~35K combined |
| Mem0 | Hybrid vector+graph, update-not-duplicate | +29.6 temporal / +23.1 multi-hop (2026 algorithm); 21 framework integrations | ~26K |
| Letta | Stateful episodic persistence, core memory + archival | Gold standard chatbot memory; 20 vector backends | ~47K |
| OpenAI Agents SDK | Context-passing + tool calls | Production SDK (replaced Swarm, Mar 2025) | N/A (closed) |
| Ruflo AgentDB | HNSW + sql.js + ONNX 384-dim | Measured 1.9×–4.7× HNSW speedup; no RL nav | ~6K |
| Benchmark | Result | Method | Grade |
|---|---|---|---|
| Mem0 LoCoMo (2026 algorithm) | 92.5 / 100, 6,956 tokens/query avg | 1,540-question long-context eval | B (vendor) |
| Mem0 LongMemEval (2026 algorithm) | 94.4 / 100, 6,787 tokens/query avg | 500-question, 6 categories | B (vendor) |
| Mem0 BEAM at 1M token scale | 64.1 / 100 | Large-scale stress test | B (vendor) |
| Mem0 BEAM at 10M token scale | 48.6 / 100 | Large-scale stress test | B (vendor) |
| Ruflo AgentDB HNSW (measured) | ~1.9× at N=20k; ~3.2×–4.7× at N=5k (recall@10 ~0.99) | Internal benchmark, ruvector NAPI | A (reproduced) |
| Princeton NLP multi-agent vs single | Single wins 64% tasks; multi adds 2.1 pp at 2× cost | Comparison benchmark, 2026 | C (secondary) |
No 2026 Grade A data for NapMem RL navigation — arXiv preprints only (Grade B until peer-reviewed).
Competitive signal: MCP has emerged as the universal plugin substrate across all 4 major competitors in Q2 2026. AWS released awslabs/agent-plugins (domain-specialized skill packages). Enterprise deployments standardizing on centralized approval models with staged rollout modes. Security audit (Grade C — single source, explicitly labelled): 12–20% of skills on public agent marketplaces found malicious — static trust-level metadata is insufficient. Ruflo's IPFS registry checks trust-level at discovery time; no behavioral smoke test before activation.
Ruflo gap: No sandboxed behavioral pre-activation test. Add fixture-request smoke test in isolated subprocess before enabling any plugin.
| Field | Value |
|---|---|
| Session commit | a444930d88d753e04793f55bd38861e82d9cb062 |
| Report SHA-256 | f48ab1c13ba2c452f0659a1635a330bf18c75d3ea57983e85b948a2a36bdd85d |
| Witness stamp | 0ed34f66403970b78e8ae70ec1b40a9e921f7ef43b2145b050a00713749847e9 |
| Verification | sha256(report_file) → concat session_commit → sha256 → must equal witness |
ADR-179 — RL-navigated memory tier for AgentDB: 3-level pyramid (hot/warm/cold) + Q-learning policy head (3 actions: shallow/deep/full). Benchmark gate: ≥5% latency reduction on shallow queries, no regression on recall@10. Continues AutoMem thread from #2536.
Upgrade AgentDB retrieval to fused scoring: α·cosine + β·BM25 + γ·entity_overlap. Benchmark against Mem0-style LoCoMo/LongMemEval eval to measure temporal and multi-hop gains.
Plugin security — sandboxed behavioral smoke test: Before enabling any plugin, run in isolated subprocess against known-safe fixture request; assert no unexpected side-effects. Pairs with existing IPFS trust-level checks.