ZenBrain: A Neuroscience-Inspired 7-Layer Memory Architecture for Autonomous AI Systems

v7 (2026-05-03): LongMemEval-500 Validation, Competitive Head-to-Head Wins & Defensive Analyses

 

NEW — LongMemEval-500 Benchmark:

- Matches long-context oracle binary-judge accuracy within 4.5 pp

  (47.7% vs. 52.2% baseline; 91.3% relative)

- Achieves this at 1/106th of the per-query token cost

- App. F.5–F.6 + Fig. 2 in paper

 

NEW — Competitive Head-to-Head (Bonferroni-corrected):

- Wins all nine head-to-head comparisons (3 competitors × 3 LLM judges)

- Compared against Letta, Mem0, A-Mem

- α = 0.05/18, p_min = 6.2×10⁻³¹, Cohen's d ∈ [0.18, 0.52]

 

NEW — Cross-Provider Judge-Bias Sanity Check:

- Δ(GPT vs. Anthropic) = -0.0001 for ZenBrain

- Δ(GPT vs. Anthropic) = -0.049 for Mem0

- Rules out LLM-judge-specific confounds

 

Cross-Session Reasoning:

- +20.7% F1 on LoCoMo cross-session dependencies

- +19.5% F1 on MemoryArena cross-session dependencies

- Multi-layer routing vs. flat single-layer baseline

 

Stress Ablation (unchanged from v6, expanded analysis):

- 9 of 15 mechanisms individually critical under stress

  (decay 0.25/day, 60 days, Wilcoxon, 10 seeds)

- ΔQ up to -93.7%

- Cooperative survival network confirmed

 

Sleep Consolidation:

- +37% stability with 47.4% storage reduction (p ≤ 5.1×10⁻³)

- TripleCopyMemory retains S(t) = 0.912 at 30 days

 

NoDecay Ablation:

- Principled forgetting costs only ΔP@5 = 0.002 (p = 0.043)

 

Honest Scope (NEW section):

- LoCoMo's substring-based aggregate F1 favors BM25 by metric design

- Explicitly acknowledged, not contested

- Advantages most pronounced on judge-graded answer quality

  and cross-session reasoning

 

Paper improvements over v6:

- Abstract reframed to lead with LongMemEval-500 + 12/12 wins

  [historical wording; these cells are correctly nine of nine — corrected in v8]

- Extended defensive analyses (Bayesian calibration, statistical

  power, failure-mode taxonomy)

- New tables for head-to-head comparison and cross-provider check

- Honest-scope paragraph added before main results

 

Stats: 47 pages, 27 tables, 3 figures, 11,589 CI tests, Phase 145, 60 tools

 

Correction (2026-08). The changelog entry above is left unchanged — it records what this version did at the time. Its count came from reading the 4 × 3 result grid as one cell per comparison, which includes ZenBrain's own column. These cells are correctly nine of nine: four systems give six pairwise comparisons, three of them involving ZenBrain, times three LLM judges. All nine clear significance under Bonferroni correction (α = 0.05/18 — the denominator is the six pairs × three judges; App. F.2). The wording of record is "wins all nine head-to-head answer-quality comparisons (3 competitors × 3 LLM judges)". Superseded by v8, which carries it.