1.2 The pitch ¶
OwnSona breaks the cycle by being the durable memory store every
MCP-capable LLM connects to.
- Tell Claude once that your dog’s name is Mochi, that you prefer concise
answers, that the quarterly forecast lives in ~/work/q3/, or
that you’re a Go developer learning React — and the next session in
ChatGPT or Gemini already knows.
- A fact written by one model is recallable by the next. Your context
follows you across providers, across models, across years.
- You own the database. It lives on a server you control. Nothing
about your memory store leaves your machine except the embedding API
calls that vectorize new facts (and even those go to a configurable
endpoint, including self-hosted models).
The mechanism is unspectacular: MCP standardizes tool-calling for
LLMs, OwnSona implements seven tools, and any LLM client that speaks
MCP can call them. Under the hood, memories are stored in PostgreSQL
with the pgvector extension; recall is vector-similarity search
over OpenAI’s text-embedding-3-small model by default
(swappable).