Core concepts
Learn the vocabulary that connects governed Assets, sources, permissions, retrieval, and delivery.
OrgMemory separates durable evidence from the derived structures that make it useful. These concepts form the vocabulary used across the product and this documentation.
Sources
Uploads, Slack, Google Drive, and GitHub adapters contribute versioned content and source-native permission evidence.
Knowledge Asset
A stable identity over immutable versions derived from the governed source ledger.
Registry Asset
A governed reusable Prompt, Work Instruction, Capability Pack, or Skill with reviews and exact releases.
Permission-aware retrieval
A query path that filters tenant, lifecycle, classification, source ACL, and OpenFGA relationships before evidence leaves the system.
Evidence and identity
Source Object is the stable record for something observed in a source system. Source Revision is immutable content for one source version. Evidence Blob records the original bytes in object storage. Query strings, fragments, and embedded credentials are removed from persisted source URIs.
An external user becomes an OrgMemory actor only through an explicit OIDC
(issuer, subject) binding. Source-native users and groups have their own typed,
stable identifiers and reach the actor model only through verified mappings.
Display names and email addresses are observations, not authorization keys.
Permissions
OrgMemory intersects multiple controls:
- OpenFGA relationships govern organization, Knowledge Space, Knowledge Asset, and registry Asset actions;
- immutable source ACL snapshots preserve what the external source allowed;
- the current ACL head applies grants and revocations without rewriting content;
- classification policy adds role and department constraints;
- the canonical ledger rechecks lifecycle, revision, and projection identity.
Passing one control never bypasses another. A product administrator is not automatically an Executive and cannot read source content solely because they manage the deployment.
Canonical data and projections
PostgreSQL stores canonical identities, revisions, ACL evidence, publication state, and audit records. MinIO stores original evidence bytes. Full-text, pgvector, graph, OpenSearch, and Neo4j structures are rebuildable projections: they accelerate retrieval but do not become permission authorities.
Two kinds of Asset
The name is intentionally shared but the lifecycles differ:
| Concept | Purpose | Ownership |
|---|---|---|
| Knowledge Asset | Searchable organizational knowledge derived from source evidence | Knowledge ledger |
| Registry Asset | A reviewed reusable capability delivered as an exact release | Asset Registry |
Capability Packs may pin both exact registry releases and exact Knowledge Asset versions. A denied component becomes an opaque access gap; its title, type, and count are not disclosed.
Retrieval and delivery
The Assistant, REST search, graph explorer, and MCP search converge on the same
permission-aware retrieval service. Registry Asset delivery separately resolves
the caller, required OAuth scope, exact release, live CAN_USE authorization,
and component-level access.
The model never authorizes a fact. It receives only the evidence closure already approved for the current request, and citations point back to authenticated API routes that recheck current access when opened.
Last updated on