Seven layers and the coordinator

ZenBrain organizes an AI agent's memory into seven layers that a coordinator orchestrates.

What happens on recall?
Working Memory Short-Term Episodic Semantic Procedural Core Cross-Context MemoryCoordinator Question
Illustration · layout invented, not a measurement

All seven layers

A click on a layer or the coordinator, in the picture or in the list, shows here what it is and what version 9 says about it. A second click shows all again.

Working Memory

Maintains the active task focus with limited capacity, about seven items. When a task is complete, items move to short-term memory.

Version 9: In the open library (version 0.5.1), it is the only layer in which anything decays.

Short-Term Memory

Holds the context of the current session and is bounded by its duration. At session boundaries, its content consolidates into the episodic and semantic layers.

Version 9: In the ablation harness, this layer exists only as a name.

Episodic Memory

Stores concrete experiences with their temporal context: what happened, when and where. Events are timestamped and support questions about when.

Version 9: In the open library (version 0.5.1), consolidation writes new facts and marks the episode it consolidated.

Semantic Memory

Contains abstracted knowledge of facts, concepts and relationships, organized as a knowledge graph; it emerges from the consolidation of episodes.

Version 9: The open library (version 0.5.1) writes a strength back only after a review that the application reports. The constants of its FSRS-style review scheduling were not fitted to review logs.

Procedural Memory

Encodes learned skills and routines: successful tool-use patterns, workflow templates, behavioral strategies. Repeated successful execution strengthens the entries.

Version 9: In the private production system, no deletion endpoint reaches this layer.

Core Memory

Holds persistent identity facts such as preferences and keeps them available in context, following the pinned memory pattern.

Version 9: An engineering layer, not a construct of memory research. In the private production system, no deletion endpoint reaches the core memory.

Cross-Context Memory

Recognizes the same entity across separate domains such as personal, work and learning, and transfers knowledge selectively between them.

Version 9: An engineering layer: merging compares entity names, and no access control gates it.

MemoryCoordinator

It orchestrates all seven layers: it routes new items to suitable layers, assembles results from several layers on recall and moves entries between layers, for example from episodic to semantic.

Version 9: In the open library (version 0.5.1), it searches the requested layers, removes near-duplicates and returns the top results. The comparisons in version 9 do not run through it but through a scoring class from the test harness and its Python port.

Per version 9: Five of the seven layers correspond to established cognitive constructs; core and cross-context are engineering layers. In the open library (version 0.5.1), decay acts only on working memory, and nothing is replayed.

Source: ZenBrain, version 9 · arXiv:2604.23878v4 · DOI 10.5281/zenodo.23832288