Published April 9, 2026
| Version v6
Preprint
Open
ZenBrain: A Neuroscience-Inspired 7-Layer Memory Architecture for Autonomous AI Systems
Description
ZenBrain v6 — Quality Evolution: Cooperative Gradient & Fair Competitive Comparison
A neuroscience-inspired 7-layer memory architecture (Working, Short-Term, Episodic, Semantic, Procedural, Core, Cross-Context) for autonomous AI systems, integrating 15 algorithms grounded in peer-reviewed neuroscience.
v6 Changes:
- NEW Challenging Ablation Suite (400 facts, decay=0.20, 50 days): 7 of 15 algorithms become individually significant, proving cooperative redundancy is genuine compensation
- 3-Level Difficulty Gradient: moderate → challenging → stress reveals four-class algorithm taxonomy (progressive, always-critical, stress-only, cooperatively redundant)
- Table 1 Extended: 9 comparison systems (added LightMem, MemoryOS, Tiwari 2026) + Neuromodulation/Reconsolidation feature rows
- Abstract & Discussion Reframed: cooperative survival network as central finding
- 95 experiment tests across 4 suites (was 77 in v5), all passing with seeded PRNG
Key Results (unchanged): 31.1x retention advantage, 84.7% emotional gap at day 60, Sleep as 1.92x multiplier, Fiedler Δ +0.051 after consolidation.
Reproducibility: All experiments use Mulberry32 seeded PRNG (10 seeds). Full source code at github.com/zensation-ai/zenbrain.
11,607 total tests | 322K LOC | Phase 145 | 60 AI tools
Notes
v6 (2026-04-09): Quality Evolution — Cooperative Gradient & Fair Competitive Comparison
Major experiment expansion across 4 suites (95 tests, was 77):
NEW — Challenging Ablation (400 facts, 50d, decay=0.20):
- 7 of 15 algorithms become individually significant (ΔQ -25.5% to -93.1%)
- Proves the moderate-condition redundancy reflects genuine cooperative compensation, not algorithm inactivity
- Establishes measurable gradient: moderate → challenging → stress
Moderate Ablation (300 facts, 45d, decay=0.15):
- Sleep impact: -34.4% (p < 0.005)
- Cooperative redundancy confirmed: no single non-Sleep algorithm removal degrades quality, yet PMA group removal causes -67.5% (p < 0.005)
- Fault-tolerance analogy: individual strands dispensable, rope essential
Stress Ablation (500 facts, 60d, decay=0.25):
- 9 Tier-1 (survival) algorithms: ΔQ -3.4% to -93.7%
- 6 Tier-2 (quality) algorithms: no individual retention impact
- Two-tier cooperative structure confirmed under extreme conditions
Competitive Comparison — Fair Baseline:
- ZenBrain and SimpleMemory share same base decay rate (0.15/day)
- SimpleMemory collapses to P@5=0 by day 30; ZenBrain retains 100%
Integration Cascade (unchanged from v5):
- 31.1x retention advantage, 84.7% emotional gap at day 60
- Sleep as 1.92x multiplier, Fiedler Δ +0.051 after consolidation
Paper improvements:
- Abstract reframed: leads with cooperative gradient, not retrieval benchmarks
- NEW Table 8: Challenging ablation with gradient evidence
- NEW Table 10: Gradient summary with four-class taxonomy
- Table 1: Extended to 9 systems (LightMem, MemoryOS, Tiwari '26)
- Discussion reframed: cooperative network as central finding
- MemoryAgentBench underperformance honestly discussed
- BM25 dominance explanation strengthened with citation
- Experimental roadmap paragraph added to guide reader
- British spellings standardized to American English
- fleming2014 bib key corrected to fleming2012
- Appendix seed count corrected (3→10)
Stats: 95 experiment tests, 11,607 total tests, 322K LOC, Phase 145, 60 tools