Knowledge Graph
HiveMind OS doesn't just answer questions — it remembers. Every conversation builds a private knowledge graph that makes future interactions smarter.
Think of it as the agent's long-term memory — a local property graph database of entities, facts, relationships, and preferences that grows with every interaction. Fully searchable. Completely private.
How Memory Works
Memory flows through three phases: observe, store, and recall.
Observe
During every conversation, HiveMind OS quietly extracts useful information — people you mention, decisions you make, technologies you discuss, preferences you express. You don't need to do anything special.
Store
Extracted knowledge goes into a local SQLite property graph — nodes and edges with rich properties. Two search layers make recall fast:
- FTS5 full-text indexing — keyword search across node names, descriptions, and observations
- sqlite-vec vector embeddings — semantic similarity search powered by a local embedding model (data never leaves your machine)
Recall
When you ask a question, the agent searches using three strategies:
- Keyword search — fast full-text matching via FTS5
- Semantic similarity — vector-based nearest-neighbour search for fuzzy, meaning-based recall
- Graph traversal — follow relationships to find connected context
HiveMind OS combines all three automatically, re-ranking by relevance, recency, and confidence.
What It Remembers
The knowledge graph uses a flexible schema of generic Nodes and Edges with string-typed labels. Rather than a fixed taxonomy, each node has a type field (a free-form string like "person", "project", "preference", "fact") and rich properties — name, description, observations, and metadata. Edges likewise have a string type (e.g. "related_to", "uses", "works_on") connecting any two nodes.
This generic structure means the graph can represent any kind of knowledge — people, projects, decisions, preferences, technologies, events — without a rigid predefined schema. Relationships connect nodes to form a traversable web of everything the agent knows.
How to Interact With It
Interacting with memory
- Teach the agent — Just ask it to remember something in natural language: "Remember that I prefer React with TypeScript and Tailwind"
- Browse your memory — Use the Knowledge Explorer in the UI to search and visualise your knowledge graph
You can ask the agent to remember anything on the spot — preferences, decisions, facts, procedures. And the Knowledge Explorer lets you browse and search your knowledge graph directly from the UI.
But most of the time, you won't need to do either. Memory is automatic. The agent extracts and stores knowledge from every conversation. Over time, it builds an increasingly detailed picture of your projects, preferences, and working patterns.
Classification and Privacy
Knowledge inherits the same classification levels as everything else in HiveMind OS. Every node carries a data class — PUBLIC, INTERNAL, CONFIDENTIAL, or RESTRICTED.
Here's the critical rule: a node inherits the highest classification of its ancestors. If a CONFIDENTIAL entity links to a PUBLIC observation, that observation's effective class becomes CONFIDENTIAL.
RESTRICTED knowledge — private keys, secrets, sensitive details — will never be included in prompts sent to PUBLIC providers. Classification is enforced at the query engine level, not as an afterthought.
See It in Action
Here's the magic moment: after a few conversations about your project — discussing architecture, picking a database, debating frontend frameworks — try asking:
"What tech stack did we decide on?"
Watch HiveMind OS recall the exact discussions, surface the decisions you made, and present a coherent answer drawn from multiple conversations. No bookmarks, no notes, no manual saving. It just knows — because it was paying attention the whole time.
Learn More
- Knowledge Management Guide — Hands-on tips for teaching your agent and managing memory
- Privacy & Security — How classification protects your knowledge
- Personas — How different agent roles access and contribute to the knowledge graph