Long-term Memory

Remember across millions of tokens, then answer from a tiny reconstructed context. Every turn is embedded into a vector store; per question the gateway assembles only the relevant slice — digest, facts, recent turns, and retrieved episodes — within a fixed token budget.