The science behind ZenBrain
9 foundational algorithms and 6 PMA components across 7 memory layers. Open-access preprint on arXiv. All documented.
“Memory is not storage — it is a living process of forgetting, consolidation, and rediscovery. We translated this process into software.”
Published research
Our technical paper is publicly available as an open-access preprint on Zenodo and arXiv, with a defensive disclosure on Elsevier TDCommons.
ZenBrain: A Neuroscience-Inspired 7-Layer Memory Architecture for Autonomous AI Systems
ZenBrain is a seven-layer, neuroscience-derived architecture for agent memory in LLM agents. An open library holds the seven layers, a MemoryCoordinator and nine foundational algorithms; six Predictive Memory Architecture components (NeuromodulatorEngine, ReconsolidationEngine, TripleCopyMemory, PriorityMap, StabilityProtector, MetacognitiveMonitor) run in a private production system. Ablating one component at a time in a simulation harness exposes an effect we call cooperative masking: under moderate load, fourteen of the fifteen ablations look costless; under heavier load, with no change to the components, ten of the fifteen move in the same direction on all ten seeds — a caution we conjecture applies beyond ZenBrain. On LongMemEval-S (500 questions), the evaluated ZenBrain arm, a Python port of a scoring class from our test harness and not the seven-layer system, answers 47.7 % of the questions correctly (95 % interval [43.3, 52.1]); with repaired adapters, Letta reaches 48.4 % and Mem0 47.2 %. It reaches 91.3 % of a full-context baseline’s accuracy (52.2 %; its answer model reads the question’s entire history) with 1/109.7 of its input tokens in the answer step; building the memory is not counted. Of nine head-to-head comparisons of the retrieved context (3 competitors × 3 LLM judges), three hold at Bonferroni-corrected significance, all against A-Mem, and the remaining six, against Letta and Mem0, show no significant difference; there the point estimates lean to the competitor. This version corrects version 8 after a post-publication audit. Open source under the @zensation npm scope, Apache-2.0.
Architecture
One architecture, many applications
ZenBrain is our open-source memory core. Multiple applications build on it — the same seven-layer architecture, only the domain adaptation varies.
Memory layers
- Working
- Short-Term
- Episodic
- Semantic
- Procedural
- Core
- Cross-Context
Counting
How we count
Three numbers from the paper count three different things: the building blocks of the architecture, the components of the ablation and the head-to-head comparisons of the benchmark. Version 9 separates them explicitly (corrections 15 and 16).
Architecture
Seven memory layers, one coordinator (MemoryCoordinator) and nine foundational algorithms in the open library @zensation/algorithms: Ebbinghaus decay, FSRS-style scheduling, Hebbian co-activation, sleep consolidation, Bayesian confidence propagation, emotional tagging, context retrieval, similarity and visualization. Six Predictive Memory Architecture (PMA) components run in the private production system. Algorithms and PMA components are different kinds of building blocks and do not add up to one number.
Source: version 9, abstract and introduction
Ablation
The ablation switches off fifteen components one at a time: the six PMA components and nine mechanisms of the evaluation harness, which run there as stand-ins at simulated values, not as production code. Three of the nine share a name with architecture modules (Hebbian, sleep, FSRS-style scheduling) but are different implementations.
Source: version 9, abstract and caption of the architecture figure
Benchmark (LongMemEval_S, 500 questions)
Of nine head-to-head comparisons of the retrieved context (three competitors, three LLM judges each), three hold at Bonferroni-corrected significance, all against A-Mem; the six against Letta and Mem0 show no significant difference. The evaluated ZenBrain arm, a Python port of a scoring class from our test harness, answers 47.7 % of the questions correctly [43.3, 52.1]. With their adapters repaired, Letta reaches 48.4 % and Mem0 47.2 %. ZenBrain thus reaches 91.3 % of the accuracy of a full-context baseline (52.2 %; its answer model reads the entire history) with 1/109.7 of its input tokens in the answer step.
Source: version 9, abstract
All three statements are in version 9 of the paper (arXiv:2604.23878v4; Zenodo 10.5281/zenodo.23832288).
Eight innovations, shipped as one system
Eight technologies shipped together in ZenAI in this combination.
Sleep Consolidation Engine
Idle-time memory consolidation, run in production in ZenAI — inspired by hippocampal replay (Stickgold & Walker 2013). Weak connections are pruned, stable ones strengthened.
Idle-time consolidation is pursued elsewhere too, for example as “dreaming” in Letta Code. The zenbrain module selects memories for replay by access frequency, emotional weight and recency.
Source: zenbrain (opens in new tab)7-Layer Memory Coordinator
Unified orchestrator that creates seven memory layers and works on five of them — from working memory to core memory. Based on Global Workspace Theory (Baars 1988).
Seven layers as classes of their own in @zensation/core, from Working Memory to Cross-Context Memory, behind one API.
Source: zenbrain (opens in new tab)A-RAG (Autonomous Retrieval Agent)
Meta-agent that creates retrieval plans before any search is executed. Heuristic-first with LLM fallback, max 4 dependent steps.
A dedicated planning layer ahead of retrieval, with a heuristic gate: simple queries never reach the planner and cost no LLM call.
Source: ZenAI snapshot (8 May 2026) (opens in new tab)Multi-Agent Debate Protocol
Structured 3-round debate protocol when agents disagree. Challenge → Response → Resolution with automatic escalation.
After three rounds without consensus the protocol escalates to a human instead of computing a majority.
Source: ZenAI snapshot (8 May 2026) (opens in new tab)Curiosity Engine
Automatic knowledge gap detection with quantified gap score. Analyzes query history, fact density, and confidence — recommends targeted actions.
Gap detection runs on its own and proposes actions rather than waiting for a question. Inspired by Loewenstein's information gap theory (1994).
Source: ZenAI snapshot (8 May 2026) (opens in new tab)Prediction Engine
User intent prediction from temporal and sequential patterns. Learns from prediction errors — the more often wrong, the better the next prediction.
Prediction runs inside the memory layer itself, from behavioural patterns.
Source: ZenAI snapshot (8 May 2026) (opens in new tab)HyperAgent L0–L2
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: self-improvement runs only inside those bounds.
Source: ZenAI snapshot (8 May 2026) (opens in new tab)Cross-Context Entity Merging
Detection and merging of entities across 4 isolated contexts (Operations, Finance, People, Strategy). Bayesian confidence updates on conflicts.
Entity identity is tracked across isolated contexts: at a similarity of 0.95 or more, entities are linked automatically; between 0.85 and 0.95 they are flagged for review.
Source: ZenAI snapshot (8 May 2026) (opens in new tab)Explore the system interactively
An accessible, interactive depiction of the memory-based system — from the 7-layer memory and its neuroscience inspiration to how the agents work together. With a toggle between plain-language and technical explanations.
The idea in one sentence: collective intelligence on a cognitive architecture — humans and machines together, on a structure that keeps knowledge. The explorer below makes exactly that architecture tangible.
These mechanisms run for real — run the open library live →
Interactive depiction · best viewed on desktop
The principle behind the quality imperative
A working system sits on a narrow edge. Change the survival rule step by step and watch the grid collapse, carry structure, or clog — live. Most rules fail; only a narrow band holds.
Why quality is not optional for us: the very efficiency that makes a system valuable amplifies every rule you give it — for better and for worse.
The full story — why quality is not optional →
Interactive · click the neighbour counts on or off · best on desktop
Seven layers of memory, grounded in neuroscience
The HiMeS architecture, inspired by the Atkinson-Shiffrin model (1968) and modern cognitive science.
Working Memory
Active focus — 7±2 items per Miller's Magical Number. Fastest access, shortest lifespan.
2026Short-Term Memory
Session context and conversation continuity. Survives the current session.
2026Episodic
Concrete experiences with emotional tagging. 400+ keyword lexicon (DE+EN) for arousal/valence scoring.
2026Semantic
Factual knowledge with FSRS-style scheduling. Spaced repetition plans when a fact is reviewed next.
2026Procedural
Workflows and skills. Tool chains are analyzed and optimized.
2026Core Memory
Immutable foundations following the Letta pattern. Pinned facts that are never forgotten.
2026Cross-Context
Knowledge and entities linked across isolated contexts — with Bayesian confidence updates on conflict.
2026Mechanisms Inspired by Neuroscience
FSRS-style Spaced Repetition
open-spaced-repetition/fsrs4ankiOptimal review timing at ~90% target retention
Ebbinghaus Forgetting Curve
Ebbinghaus (1885)Exponential decay with configurable half-life
Hebbian Learning
Hebb (1949)Co-activated facts strengthen connections (×1.09/activation)
Homeostatic Normalization
Turrigiano (2004)Prevents runaway growth of edge weights
Sleep Consolidation
Stickgold & Walker (2013)Hippocampal replay with +50% stability boost
Synaptic Homeostasis
Tononi & Cirelli (2006)Weak connections pruned during sleep
Emotional Modulation
LeDoux (1996)Emotional memories decay 2.7× slower
Bayesian Propagation
Pearl (1988)Confidence updates across the entire knowledge graph
Global Workspace Theory
Baars (1988)Conscious access through competitive context assembly
Information Gain Scoring
Shannon (1948)Entropy-based prioritization of new facts
Knowledge Gap Theory
Loewenstein (1994)Systematic detection of missing knowledge
Working Memory Capacity
Miller (1956)7±2 active items in working memory
Neuromodulation (PMA)
Schultz (1997) · Aston-Jones (2005)Dopamine, NE, 5-HT, ACh — four channels with tonic + phasic dynamics
Reconsolidation (PMA)
Nader (2000) · Schiller (2010)Memory destabilizes on retrieval — four PE-gated update modes with rollback
Triple-Copy Memory (PMA)
Squire & Bayley (2007)Three traces with divergent dynamics: fast (4h), medium (14d), deep (logarithmic)
Intelligent knowledge retrieval
6 strategies, dynamically selected per query. Not one pipeline — an adaptive system.
End-to-end RAG pipeline
↻ Self-RAG Critique: loops back if confidence < 0.5
A-RAG Planning
Meta-agent plans retrieval steps before execution. Heuristic-first, LLM fallback.
GraphRAG 3-Layer
Event subgraph + semantic graph + community summaries. 5 parallel strategies.
Self-RAG Critique
Automatic reformulation at confidence < 0.5. 4-component scoring.
HyDE Retrieval
Hypothetical answer → embedding → search. Auto-detection with 5s timeout.
Contextual Retrieval
Chunk enrichment per Anthropic method. +67% retrieval accuracy.
Embedding Drift Detection
BullMQ worker monitors drift >10%. Automatic cache invalidation.
Autonomous agents with safety bounds
Multi-agent orchestration with structured debate, dynamic team composition, and recursive self-improvement.
Multi-agent team architecture
3 rounds: challenge → response → resolve
Debate Protocol
3-round debate on disagreement. Challenge → Response → Resolution.
Dynamic Team Builder
5 specialist agents, automatically composed by task type.
HyperAgent L0–L2
Recursive self-improvement with daily budgets, sandbox tests, and auto-rollback.
Persistent Agent Loops
Pause/resume with state checkpointing. Long-running tasks over days.
A2A Protocol
Agent-to-agent communication per Google standard. /.well-known/agent.json discovery.
Implicit Feedback
Automatic behavior detection without explicit labeling.
AI that thinks about thinking
Curiosity, prediction, metacognition — three pillars of cognitive intelligence.
Curiosity Engine
Quantified gap score: query frequency × fact density × confidence × RAG quality.
Prediction Engine
Temporal + sequential pattern recognition. Learns from prediction errors.
Metacognition
Confidence calibration, confusion detection, capability profiling.
Adaptive Thinking
4-tier thinking budgets: 1K→16K→64K→128K tokens. Auto-detection.
91.3% of the accuracy with 1/109.7 of the input tokens in the answer step
On LongMemEval-500, the measured ZenBrain arm (per the correction notice a Python port of a test-harness class, not the seven-layer system) reaches 91.3% of the accuracy of a full-context comparison arm (whose answer model reads the whole history) with 1/109.7 of its input tokens in the answer step.
In our own measurements on LongMemEval-500, three of nine pairwise comparisons hold, all three against A-Mem, significant under Bonferroni correction, across three independent LLM judges. The six against Letta and Mem0 show no significant difference, judged at equal evidence depth and under version-matched judges.
Open and reproducible
ZenBrain is open source. All algorithms, all tests, all documentation — openly available.
Frequently asked questions about the architecture
What makes ZenBrain different from a vector store?
A vector store searches for similarity in a flat index. ZenBrain models memory as a process: what is used repeatedly consolidates, what is left alone becomes harder to recall. Seven layers separate working memory, episodic, semantic and procedural knowledge, and a consolidation phase at rest condenses experience into knowledge. The difference shows where context has to hold for weeks rather than for one conversation.
Why seven layers instead of one?
Because different kinds of memory need different rules. Working memory has to be fast and volatile, semantic knowledge durable and condensed, procedural memory retrievable without deliberate search. Each layer carries its own decay and consolidation logic; on top of that come nine foundational algorithms and six components of a predictive architecture.
Are the algorithms actually grounded in neuroscience?
They draw on memory research, not on an analogy, and the paper names the sources: among them FSRS for spacing intervals, Hebbian learning, the Ebbinghaus forgetting curve, Bayesian confidence propagation and sleep consolidation. The two-factor synaptic model and the vmPFC prediction-error coupling go back to neuroscience work (Iatropoulos, Gerstner and Brea 2025; Zou et al. 2025); the concrete form — weight plus consolidation variance, the prediction-error signal — is our own construction.
How does ZenBrain compare to other open memory systems?
On LongMemEval-500, at equal evidence depth and judged by three independent LLM judges, three of nine head-to-head comparisons hold for ZenBrain, all three against A-Mem. The remaining six, against Letta and Mem0, show no significant difference. All nine contrasts are Bonferroni-corrected and judged under version-matched judges. The measured ZenBrain arm (per the correction notice a Python port of a test-harness class, not the seven-layer system) reaches 91.3 percent of the accuracy of a full-context comparison arm (whose answer model reads the whole history) with 1/109.7 of its input tokens in the answer step.
Can I verify the results myself?
Yes, and that is the point. The memory core is open under Apache 2.0, the publications carry DOIs on Zenodo, and the benchmark page gives every figure its method, source and verification path. Anyone who doubts a number should be able to recompute it without asking us.
Is the paper peer-reviewed?
The preprint itself (arXiv:2604.23878) is not peer-reviewed. The neuroscience it builds on is. We state that plainly, because the distinction matters — and because open evidence has to carry a claim, not a reference to a process.
More from this research
Related pages
Core and application fields, on a shared ethics foundation.
- Research Three tracks: cognitive architectures, safety in public spaces, applied AI.
- Adopt AI A neutral roadmap for AI adoption — prioritisation, stage-gates, KPIs and EU AI Act / GDPR governance, with an interactive simulation.
- Methodology Pre-registration, reproducibility, external validation, data minimisation — operational standards in detail.
- Publications arXiv, Zenodo (DOI), software releases, open-access principles.
- Benchmarks ZenBrain by the numbers: every figure with its method, source and verification link.
- Resources Code, replication material, BibTeX citation, licences, identifiers.
- Research ethics GDPR Art. 89, AI Act Art. 5, Brokdorf line. Eight mandatory corrections.
- Public sector & funding Research offerings for BMBF, BBK, universities and research consortia.
- Principal Investigator Profile, background, identifiers, contact paths.