A growing memory store was going to slow the assistant down,
so we built it sparse from day one
An AI assistant's memory needed to grow with usage without every lookup getting slower.
Task
A memory engine that stores every conversation in full eventually gets slow to search and expensive to keep. SENET needed a memory representation that stays fast to query, no matter how much the assistant has seen.
What we did
We built the memory engine on a sparse distributed representation. Each memory is a 4096-dimension vector with only 82 bits active at once. That sparsity keeps comparisons cheap even as the store grows. A two-level memory structure also searches recent context and long-term memory differently.
Result
The engine runs inside the assistant’s FastAPI backend, with no separate vector database to operate. The router built on top of it reached 0.91 accuracy choosing between memory modes. The design is part of what makes SENET a product a buyer can run themselves, not a service that depends on us.
Full story: SENET: a sellable AI assistant with its own memory engine and a self-training router. Want the same? Get a plan.