One persistent memory can serve several models. XAB coordinates reasoning by role or task, using local open models or remote services via API. AI consoles can access the memory through MCP.
Persistent memoryKnowledge, relationships & history
PostgreSQL + pgvectorNeo4jVersioned archive
↕
OrchestrationContext & coordination
Model reasoningLocal, open models or models via API
Console accessAI consoles through MCP tools
Model reasoning and MCP access have separate roles. Sources, models and access rules are configured for each instance.
01Storage and orchestration
Give each layer a clear job.
Persistent storage, graphs, retrieval and scheduled processing keep knowledge available beyond a model’s context window. Connectors bring in information, the orchestrator coordinates tools and model calls, and the records remain in the domain’s infrastructure.
PostgreSQL
Stores records and relationships, including the provenance and temporal references that give them meaning.
pgvector
Supports semantic retrieval of passages related to the question being investigated.
Neo4j
Makes entities and their relationships navigable as a graph.
Versioned archive
Preserves original documents and their revisions. Retrieved passages retain a reference to their source and version.
02Model-agnostic memory
Choose the model for the job.
Memory is independent of any single model. Several models can draw on the same persistent layer, with different models assigned to different roles or tasks.
Local open models
Open models can run in a local runtime. The model performs inference while XAB supplies relevant context and preserves memory independently of that runtime.
Remote models via API
Models can also run as external services accessed through their APIs. Their processing and deployment locations depend on the service selected.
MCP connects the client to the memory.
An MCP-compatible AI console or client can connect to XAB’s memory and tools. MCP is the access protocol; inference is performed by the selected local runtime or remote API service.
03Dual graph and relationship matrix
Things have histories.
The entity graph describes people, organisations, projects, places and concepts. The episodic graph connects events, decisions, constraints and changes. A relationship matrix links identities and records across memory layers, carrying provenance and temporal references so a model can find the original evidence.
Meaning over time
A lasting description, a temporary state, an event and an infrastructure location have different meanings. A role assignment can expire or be replaced; the event that recorded the change remains part of the history.
Connected layers
The matrix connects source material, documents, conversations, diaries and activity. The infrastructure map identifies where relevant services and data are located.
04Retrieval
Follow a question through the evidence.
GraphRAG uses entities and relationships to narrow the search. Semantic retrieval finds relevant passages. Both work with the versioned archive, so a retrieved fragment can be checked in its original context.
Orient
Use the matrix to identify the people, projects, events and documents connected to the question.
Retrieve
Follow graph relationships, recover context from diaries and conversations, and find passages through semantic search.
Verify
Read original sources, compare dates and revisions, and make disagreements between records visible.
Compose
Bring the evidence into a response that distinguishes documented information from what remains uncertain.
05Salience, rhythm and geometry
What matters changes with the work.
Adaptive decay, inspired by the Ebbinghaus curve, reduces the prominence of information that has not been recalled. Repeated use slows that decay. Documents, episodes and relationships remain available when attention moves elsewhere.
Heat and weight
Heat reflects recent use. Weight reflects recall frequency, graph centrality and the demands of work in progress.
Kairos
Session rhythm, pauses and completed activity indicate when to explore further or consolidate. Kairos distinguishes the opening, flow and consolidation phases of a session.
Poincaré, in 16 dimensions
A hyperbolic representation places structurally related records near one another, including records without a direct link. Proximity suggests candidates to explore; relationships, documents and evidence are then checked.
06Curation and scheduled processing
How a memory was made matters.
Origin and processing determine how a record enters the memory. Curated and consolidated records enter the graph; raw material stays outside it, with its provenance preserved.
Raw
Automatically acquired information or an inference that has not been consolidated.
Curated
An intentional human record or explicit decision, retaining its author, source and context.
Consolidated
The result of automated contextual selection, reformulation, entity resolution and evidence linking.
Between sessions
Scheduled cron jobs run triage, consolidation and synchronisation; recalculate salience and Poincaré positions; update records; and check integrity, provenance and consistency across archive and graph.
07Development snapshot · October 2026
From domain memory to a network.
Each instance maintains its own ontology and access boundaries. Federation develops the ways those memories can contribute to a wider picture.
V1
Domain instances send summaries, decisions and blockers to an aggregate memory through defined exchanges.
V2 direction
Horizontal and multilevel networks, with explicit rules for what is shared and who can access it.
What could a shared memory make possible?
Explore how cognitive infrastructure can support communities of knowledge and data cooperatives.
