Research project · Memory & context
Synapse
How should an AI system remember - and how does the person stay in control of what it remembers?
In progressSynapse is our research into persistent memory. It is the shared memory layer behind our other projects: AI Workforce agents need to know what another agent decided last month, and Alvi needs to know who it is talking to.
What we're investigating
Structure over transcripts
Turning raw conversation into memory entities - facts, preferences, goals and projects - and testing whether that reduces hallucination and improves coherence across sessions.
Hybrid retrieval
Combining vector search for semantic recall with a knowledge graph for relationship-aware, multi-hop and time-aware retrieval, compared against plain retrieval-augmented generation.
Memory the user can see and edit
Every memory entry inspectable, editable, mergeable and deletable, with nothing stored without explicit consent, and local-first or EU-hosted encrypted storage.
Approach
- 1Extract structured entities from interactions, with consent.
- 2Store them in both a vector index and a graph.
- 3Choose what to inject into each prompt, and measure what that choice costs and gains.
- 4Give the user full visibility of, and control over, everything held.
Where it stands
In active development as the memory layer for our other research systems.
Working on the same problem?
We'd like to hear from researchers, engineers and teams who have run into it too.
Get in touch