Skip to main content

Memory Fabric

GrayMatter is the durable memory, control, and governance layer for deployable AI systems. It gives agents a governed way to save, retrieve, and reason over business context without scattering embeddings across every table.

Product Thesis​

GrayMatter is not a generic chatbot memory store. It is an agentic memory fabric over a customer-isolated relational object graph:

  • business objects stay normalized
  • semantic lookup is centralized
  • ACL/RBAC boundaries remain authoritative
  • provenance is preserved
  • retrieval, compression, and reindexing can become measurable operations

Central Semantic Index​

The first executable slice uses a central semantic index. Each index row references a source object rather than duplicating the object as ungoverned text.

Core reference fields:

  • targetType
  • targetId
  • optional owner/principal scope
  • optional organization scope
  • optional linked MemoryEntry
  • tenant/deployment partition metadata

Embedding metadata is stored as its own concern:

  • source text hash
  • provider and model
  • dimensions
  • vector bytes or encoded representation
  • checksum
  • compression strategy
  • status and timestamps

Agents query the semantic index, receive object references, then lazy-load source objects and graph neighbors through normal ThorAPI services.

Why This Matters​

This design avoids three common memory failures:

  • embedding columns added ad hoc to every business table
  • bypassing generated RBAC with a separate vector service
  • losing provenance when memory is copied into opaque chunks

GrayMatter keeps memory useful because it remains attached to the same object graph that governs the application.

Related: