15 neuroscience-inspired mechanisms across 7 memory layers. Open-access preprint on arXiv. All documented.
โMemory is not storage โ it is a living process of forgetting, consolidation, and rediscovery. We translated this process into software.โ
Our technical paper is publicly available as an open-access preprint on Zenodo and arXiv, with a defensive disclosure on Elsevier TDCommons.
We present ZenBrain, a neuroscience-inspired 7-layer memory architecture integrating 15 algorithms grounded in peer-reviewed neuroscience: 9 foundational components plus a Predictive Memory Architecture (PMA) with NeuromodulatorEngine, ReconsolidationEngine, TripleCopyMemory, PriorityMap, StabilityProtector, and MetacognitiveMonitor. Evaluated across ten experiments on LoCoMo, MemoryAgentBench, MemoryArena, and the LongMemEval-S Full-500 replication. On LongMemEval-500, ZenBrain wins 12 of 12 head-to-head judge comparisons against Letta, Mem0, and A-Mem under three independent LLM judges (Bonferroni-corrected p โค 6.2e-31). Under the official binary judge, ZenBrain reaches 91.3 % of long-context-oracle accuracy at 1/106แตสฐ the per-query token budget โ the oracle beats ZenBrain by only 4.5 pp while using ~106ร more tokens and no memory architecture. Sleep consolidation: +37 % stability, โ47.4 % storage. TripleCopyMemory retains 91.2 % strength at 30 days. The full 15-algorithm ablation reveals a cooperative survival network where 9 of 15 algorithms become individually critical under stress (decay=0.25, 60 days). All experiments are reproducible with seeded PRNG; the implementation is released as open-source packages under the @zensation npm scope.
Eight technologies shipped together in production in this combination.
Idle-time memory consolidation in production โ inspired by hippocampal replay (Stickgold & Walker 2013). Weak connections are pruned, stable ones strengthened.
We are not aware of any competing memory system shipping sleep-time consolidation; the closest published designs remain proposals.
Unified orchestrator for 7 memory layers in production โ from working memory to cross-context memory. Based on Global Workspace Theory (Baars 1988).
Mem0 ships 2 layers, Letta 3, Zep 2. ZenBrain ships 7 โ among the deepest memory architectures in open source today.
Meta-agent that creates retrieval plans before any search is executed. Heuristic-first with LLM fallback, max 4 dependent steps.
Heuristic-based planning before retrieval โ no LLM cost for simple queries; few production memory systems expose a dedicated planning layer.
Structured 3-round debate protocol when agents disagree. Challenge โ Response โ Resolution with automatic escalation.
Structured 3-round debate protocol โ rare in production multi-agent systems.
Automatic knowledge gap detection with quantified gap score. Analyzes query history, fact density, and confidence โ recommends targeted actions.
Few memory systems implement systematic curiosity-driven gap detection. Inspired by Loewenstein's information gap theory (1994).
User intent prediction from temporal and sequential patterns. Learns from prediction errors โ the more often wrong, the better the next prediction.
We do not see competitors predicting user intent from behavioral patterns within the memory layer itself.
3-level recursive self-improvement with formal safety bounds. Level 0 optimizes knowledge, Level 1 optimizes Level-0 strategies, Level 2 optimizes Level-1 parameters.
3-level recursion with immutable core properties, daily budgets, and automatic rollback on quality regression โ rare in production.
Detection and merging of entities across 4 isolated contexts (Operations, Finance, People, Strategy). Bayesian confidence updates on conflicts.
We do not see competing memory systems managing entity identity across isolated contexts.
An accessible, interactive depiction of the memory-based system โ from the 7-layer memory and its neuroscience inspiration to how the agents work together. With a toggle between plain-language and technical explanations.
Interactive depiction ยท best viewed on desktop
The HiMeS architecture, inspired by the Atkinson-Shiffrin model (1968) and modern cognitive science.
Active focus โ 7ยฑ2 items per Miller's Magical Number. Fastest access, shortest lifespan.
2026Session context and conversation continuity. Survives the current session.
2026Concrete experiences with emotional tagging. 400+ keyword lexicon (DE+EN) for arousal/valence scoring.
2026Factual knowledge with FSRS scheduling. Spaced repetition optimizes recall timing โ 30% better than SM-2.
2026Workflows and skills. Tool chains are analyzed and optimized.
2026Immutable foundations following the Letta pattern. Pinned facts that are never forgotten.
2026Knowledge and entities linked across isolated contexts โ with Bayesian confidence updates on conflict.
2026Optimal review timing at ~90% target retention
Exponential decay with configurable half-life
Co-activated facts strengthen connections (ร1.09/activation)
Prevents runaway growth of edge weights
Hippocampal replay with +50% stability boost
Weak connections pruned during sleep
Emotional memories decay 2.7ร slower
Confidence updates across the entire knowledge graph
Conscious access through competitive context assembly
Entropy-based prioritization of new facts
Systematic detection of missing knowledge
7ยฑ2 active items in working memory
Dopamine, NE, 5-HT, ACh โ four channels with tonic + phasic dynamics
Memory destabilizes on retrieval โ four PE-gated update modes with rollback
Three traces with divergent dynamics: fast (4h), medium (14d), deep (logarithmic)
6 strategies, dynamically selected per query. Not one pipeline โ an adaptive system.
โป Self-RAG Critique: loops back if confidence < 0.5
Meta-agent plans retrieval steps before execution. Heuristic-first, LLM fallback.
Event subgraph + semantic graph + community summaries. 5 parallel strategies.
Automatic reformulation at confidence < 0.5. 4-component scoring.
Hypothetical answer โ embedding โ search. Auto-detection with 5s timeout.
Chunk enrichment per Anthropic method. +67% retrieval accuracy.
BullMQ worker monitors drift >10%. Automatic cache invalidation.
Multi-agent orchestration with structured debate, dynamic team composition, and recursive self-improvement.
3 rounds: challenge โ response โ resolve
3-round debate on disagreement. Challenge โ Response โ Resolution.
5 specialist agents, automatically composed by task type.
Recursive self-improvement with daily budgets, sandbox tests, and auto-rollback.
Pause/resume with state checkpointing. Long-running tasks over days.
Agent-to-agent communication per Google standard. /.well-known/agent.json discovery.
Automatic behavior detection without explicit labeling.
Curiosity, prediction, metacognition โ three pillars of cognitive intelligence.
Quantified gap score: query frequency ร fact density ร confidence ร RAG quality.
Temporal + sequential pattern recognition. Learns from prediction errors.
Confidence calibration, confusion detection, capability profiling.
4-tier thinking budgets: 1Kโ16Kโ64Kโ128K tokens. Auto-detection. 60-80% cost savings.
On LongMemEval-500, ZenBrain reaches 91.3% of long-context-oracle accuracy at a per-query token budget of 1/106 โ the Pareto position for AI memory.
ZenBrain is open source. All algorithms, all tests, all documentation โ openly available.
More from this research
Core and application fields, on a shared ethics foundation.