Cerberus
An AI assistant with a long memory, built on a personal knowledge graph.
Extracts entities, decisions and preferences from everyday conversations into a persistent knowledge graph, then retrieves from it to recall context across multi-week conversations and answer questions about time.
- RAG
- Knowledge graph
- Embeddings
- LLM integration
The problem
Chat assistants forget. Every conversation starts from zero, so you re-explain the same context, decisions and preferences again and again, and questions like what did I decide about this last month? have nothing to draw on.
Approach
Cerberus keeps a persistent personal knowledge graph. As you talk to it, it extracts entities, decisions and preferences and stores them as structured, connected facts.
When you ask something, a retrieval-augmented generation (RAG) pipeline combines embeddings with the graph to pull in the history that matters for the task at hand, including when things happened.
How it works
Conversation
Everyday interactions with the assistant.
Extraction
Entities, decisions and preferences pulled out of each exchange.
Knowledge graph
Connected, persistent facts that build up over weeks.
Retrieval
Embeddings and graph context select what's relevant to the current task.
Answer
A response grounded in recalled history, including when things happened.
Key decisions
A graph, not a transcript
Storing structured entities and decisions instead of replaying raw chat logs makes memory queryable. Relationships and time become first-class, so temporal questions get real answers.
Retrieval driven by the current task
Context is chosen from the user's current task and interaction history, so recall stays relevant instead of flooding the prompt with everything ever said.
Results
Context recall across multi-week conversations, and answers to temporal questions such as what was decided, and when.