Welcome to OrgMemory

OrgMemory is a self-hosted, governed memory layer for enterprise AI. It turns uploads and approved connector evidence into versioned, citable knowledge while preserving source permissions across search, Assistant, API, and MCP delivery.

Current maturity

OrgMemory is a production-shaped proof of concept. Its architecture, automated security checks, deployment pipeline, and rollback path are real; pilot readiness still depends on the operational evidence listed under Limitations.

The problem

Enterprise AI systems often copy content into an index and lose the context that made the content trustworthy: which source revision was read, who could read it, which model created a projection, and which evidence supported an answer. OrgMemory keeps those facts inside the retrieval boundary.

It is designed for three related jobs:

preserve source evidence as immutable revisions rather than mutable blobs;
make authorization part of candidate selection, ranking, graph expansion, citations, and Asset delivery;
package reusable organizational capability as governed releases instead of unreviewed prompts or files.

What is available

Governed knowledge

Manual uploads and supported connectors enter one immutable source ledger with permission evidence.

Secure retrieval

Keyword, vector, and graph retrieval operate only over the current actor’s authorized evidence scope.

Reusable Assets

Prompt Templates, Work Instructions, Capability Packs, and Skills move through draft, review, release, and controlled delivery.

Multiple delivery surfaces

The web product, REST API, Assistant, MCP server, and CLI reuse canonical authorization and audit boundaries.

Choose your path

Evaluate the product: start with the system description, then review coverage and limitations.
Run it locally: follow the POC quickstart.
Administer a deployment: read identity and permissions and sources and connections.
Integrate an AI client: use the Assistant and MCP guide.

Scope boundaries

OrgMemory is not a source-of-truth replacement, a public content marketplace, or an authorization system inferred from email addresses. It does not grant access because an AI model found a relevant chunk. Source systems remain authoritative for their content permissions, OpenFGA governs product relationships, and the canonical ledger decides what may enter a response.