Track A

Memory architectures for autonomous AI systems

„How can AI remember, reason, and learn over time — without losing efficiency, reliability, or methodological transparency along the way?"

A seven-layer memory architecture grounded in neuroscience and in use in an application of our own.

Motivation

Why this research

Today's large language models forget. Their context windows are technically bounded, and plain vector stores are not enough to faithfully model the layers of human memory — from a fleeting working memory to consolidated long-term knowledge.

What is missing are architectures that explicitly model consolidation, directed forgetting, confidence propagation, and predictive memory formation. This research answers with an architecture that did not emerge from marketing but from the question of how a system can maintain its own knowledge base to an academic standard.

Methodological approach

Seven layers, nine algorithms, one coordinator

The architecture splits memory into seven functionally distinct layers — from working memory for active tasks, through episodic memory for concrete experiences, to cross-context memory for connections that span layers. Each layer is addressable through a unified API; a MemoryCoordinator orchestrates storing, retrieving, consolidating, and forgetting.

Mechanisms drawn from memory research provide the mechanics, among them FSRS-style scheduling for directed spaced repetition, Hebbian learning for co-activation, the Ebbinghaus forgetting curve, Bayesian confidence propagation, a two-factor synaptic model and a vmPFC-coupled prediction-error update; the paper counts nine foundational algorithms and six PMA components. Sleep consolidation simulates offline memory replay.

↑↓1Working2Short-Term3Episodic4Semantic5Procedural6Core7Cross-Context
The seven memory layers of the ZenBrain architecture, orchestrated by a MemoryCoordinator.

Empirical result

What the architecture shows in evaluation

3 · 6 · 0hold against A-Mem · no significant difference against Letta and Mem0 · losses
91.3%of the accuracy of a comparison arm that reads the whole history, with 1/109.7 of its input tokens in the answer step
3 judgesversion-matched, Bonferroni-corrected

Measured on a Python port of a scoring class from our test harness, not on the seven-layer system; building the memory is not counted in the 1/109.7. From the open-access preprint (arXiv 2604.23878) as amended by the correction notice (10.5281/zenodo.22831086): LongMemEval-500 benchmark, three independent LLM judges (Sonnet 4.5, Opus 4.6, GPT-4o).

Full results and methodology

Core components

The architecture at a glance

7 memory layers

Working, Short-Term, Episodic, Semantic, Procedural, Core, and Cross-Context Memory — each with its own time scale, persistence, and consolidation logic.

MemoryCoordinator

One API for the layers: store(), recall(), consolidate(), decay(), getReviewQueue(). On store, routing picks the layer by content, emotional weight and confidence.

FSRS-style spaced repetition

Free Spaced Repetition Scheduler — vmPFC-coupled variant with prediction-error update. Facts are reviewed at optimal moments, not on a fixed grid.

Sleep consolidation

Offline memory replay as a methodological core. Selection of important memories for reinforcement; directed pruning of weak associations.

Two-factor synaptic model

Consolidation is driven by two factors — co-activation frequency and a neuromodulator signal. Reproduces experimental findings from the synaptic-tagging literature.

Predictive memory architecture

Six components — NeuromodulatorEngine, ReconsolidationEngine, TripleCopyMemory, PriorityMap, StabilityProtector, MetacognitiveMonitor — model predictive memory formation with bias detection.

Get involved

Collaboration on this research

Enquiries from universities, research consortia or funding programmes — on replication studies, consortium proposals or methodological discussion — are explicitly welcome.

research@zensation.ai