Skip to main content
← Back to Help Center

How Unimatrix Works

Follow context from a supported AI tool into organized memory, through evidence-aware retrieval, and back into a new conversation. Each layer serves a purpose: continuity, relevance, inspectable corrections, or access control.

The 7-Step Memory Pipeline

When an AI client or user sends context to Unimatrix, it passes through 7 distinct pipeline stages:

1. CAPTURE — Ingest via MCP, Browser Extension, REST, Tool Calls, or Mobile.
2. UNDERSTAND — Normalize and extract atomic fact triples.
3. ORGANIZE — Categorize into canonical Spaces → Locations.
4. RESOLVE — Cross-check against Veritas graph for contradictions.
5. ENCRYPT & STORE — Persist with AES-256-GCM encryption at rest.
6. RETRIEVE — Run hybrid vector (pgvector) + keyword (tsvector) search.
7. DELIVER — Format and inject relevant memory payload into the AI client.

1. Capture

Context enters Unimatrix explicitly. In MCP clients (Claude Desktop, Cursor, Windsurf, Continue, VS Code), AI assistants invoke tools like unimatrix_store_memory. In web LLMs, the Manifest V3 browser extension captures highlighted text. Custom agents use the Fastify REST endpoints.

2. Understand

Raw transcripts are processed by the Librarian Intelligence Engine. It extracts atomic semantic triples (Subject → Predicate → Object), generates summaries, and calculates confidence scores. Local HuggingFace ONNX embeddings (BGE-small) create 384-dimensional vector representations.

3. Organize

Memories are organized in Spaces → Locations → Memories so you can scope recall to the right project. Authorization and database policies enforce access; the hierarchy makes context easier to navigate.

4. Resolve

Veritas tracks support, opposition, and uncertainty alongside provenance and supersession history. Corrections remain traceable, and total conflicts stay marked for review. Provenance aggregation deduplicates repeated memory IDs and already-counted source roots when lineage is known.

5. Encrypt & Store

Memory content is encrypted at rest using application-layer AES-256-GCM with per-record HKDF key derivation and stored in PostgreSQL with pgvector. Dashboard composer memories can optionally be encrypted client-side in your browser.

6. Retrieve

Recall can combine compressed semantic candidates, pgvector search, full-text keywords, and graph relationships. Fusion and full-precision reranking refine relevance. ARTD and graph-aware utility adjust retention using evidence, time, and use. Configured retention expiry keeps utility at its floor, while bounded utility weighting keeps a bonus from inventing relevance.

7. Deliver

The formatted memory payload is delivered back to the requesting AI client via MCP streamable-http, extension overlay, or REST API response.

Glossary of Technical Terms

AMG — Associative Memory Graph

The 2.0 server links indexed memories and supports explicit spreading-activation recall over related context. Co-activation can reinforce relationships; worker consolidation links higher-level memories to their sources. Availability depends on the deployed services and migrations.

CSMTER — Compressed Semantic Memory with Tiered Error Recovery

A layered retrieval design using compressed semantic candidates, keyword search, full-precision reranking, and structured facts. Individual request paths use different combinations of these stages.

ARTD — Adaptive Reliability-weighted Temporal Decay

An evidence-aware decay model. Corroboration, contradiction, time, and usage affect retention; graph connectivity contributes to utility. These are ranking signals, not probabilities that a statement is true.

CWBF — Confidence-Weighted Belief Fusion

Retrieval fusion that combines strategy relevance with confidence and graph-belief signals, including whether a memory has been superseded.

CDSC — Content-Defined Semantic Chunking

Splits long content around changes in meaning, giving retrieval coherent units rather than relying only on fixed character boundaries.

Memory

One saved piece of context — a fact, a decision, a preference, a snippet of a conversation. The atomic unit of memory in Unimatrix.

Space

A top-level container that groups related memories, usually one per project, organization, or domain.

Location

A themed directory or semantic container inside a Space (e.g., "Architecture Decisions", "API Specifications").

Context

Everything an AI needs to know to be useful in a session. Unimatrix dynamically retrieves relevant context so you do not have to copy-paste background information.

The Librarian / Intelligence Engine

The background process that normalizes, decomposes, categorizes, and cross-checks memories against past beliefs.

Veritas Lineage Graph

The Dempster-Shafer belief tracking graph that tracks evidence for/against facts and preserves historical supersession chains.

Embedding

A numeric vector representation of text meaning. Unimatrix uses local HuggingFace ONNX embeddings (BGE-small) for fast vector retrieval in pgvector.

MCP

Model Context Protocol — an open standard enabling AI clients (Claude Desktop, Cursor, Windsurf, Continue) to invoke memory tools directly.