How the knowledge system fits together
Documents, conversations and connected systems
Content and its source context enter through configured integrations.
Ingestion and knowledge extraction
Prepare searchable text, embeddings, metadata, entities and relationships.
Vector index
Qdrant / embeddings
Find related content by meaning.
PostgreSQL
Source metadata, projects and timestamps. Keyword search.
Neo4j + Graphiti
Entities, connected facts and their temporal context.
Hybrid retrieval and orchestration
Interpret the question, search the relevant stores, combine and rank results, then assemble context for the response.
Applications, agents and the dashboard
Use retrieved knowledge with its supporting sources.
Evidence → specialist assessments → decision case → human review
Domain rules and calculations support the assessment. The general Decision Engine and reusable outcome feedback remain in development.
Engram needs more than a list of similar documents to answer questions about an organisation. Its documented architecture combines a vector index for semantic matches, a relational database for structured records and a graph for relationships over time. These representations answer different parts of a question.
The architecture names Qdrant for vector search, PostgreSQL for records and keyword search, and Neo4j with Graphiti for temporal knowledge. Ingestion prepares the content and extracts the information those stores need. The retrieval layer searches the relevant stores and combines their results into context that an application or agent can use.
For example, a question about a delayed shipment may need the original update, the affected supplier and the orders connected to it. Similar wording helps find the update. Structured records identify the shipment, while relationships connect it to the wider business impact. The answer depends on those connections having been captured correctly.
Retrieval is a starting point
Embeddings provide a way to retrieve semantically related content. They are useful for locating material even when a question uses different wording from the source.
Engram’s documented architecture also preserves facts and relationships. Applications can use this structured context alongside retrieved text instead of treating a single search result as a complete account.
Two questions about time
When was a fact valid, and when did the system learn it? These are different questions. A late-arriving correction may change your understanding of an earlier event.
The temporal knowledge architecture distinguishes those times and preserves superseded statements. This is documented design; the precise available behaviour needs to be demonstrated in the domain implementation.
The decision process
Hard constraints and calculations should remain explicit. A focused model judgment can assess a bounded question while code controls what happens with its result.
The supply-chain pilot has a decision ledger and council process. Specialist assessments contribute to a decision case for human review. This is a domain workflow that uses the knowledge, rather than another kind of database.
The general Decision Engine and reusable outcome feedback remain in development. The diagram separates that work from the knowledge and retrieval architecture; it does not present a complete autonomous decision system as available today.
Architecture and deployment are different questions
The diagram explains component responsibilities. It is not a claim that every Engram installation runs the same services or has every connector enabled. The vault also describes an Engram-OSS memory design with session context, persistent memory and graph recall. Those layers should not be treated as an identical deployment of the organisational knowledge system shown here.
For a particular deployment, check which sources are connected, how access is restricted and whether retrieval preserves the source and time context you need. Extracted relationships and summaries can be wrong. Keeping the underlying evidence available gives the reviewer a way to challenge them.