Adapt: open-source memory that rewrites its own structure when your data surprises it

TypeScript library I've been building. The thing I want feedback on: when the memory keeps dismissing data it shouldn't, or queries keep coming back empty, it restructures itself;splits overloaded learners, merges redundant ones, grows new ones for domains nobody configured. Works with any Vercel AI SDK provider, so Ollama / LM Studio / llama.cpp plus hosted. Observe and synthesize are separate LLM calls, so you can run a 7B for observe (the high-volume path) and a bigger model for synthesize. Cuts cost a lot vs one big model for everything. MIT, SQLite persistence, no vector DB required.