FAQ
Where does my data live?
Section titled “Where does my data live?”In your infrastructure. The memory store is either a directory on disk (file backend) or your own PostgreSQL database. The engine, the MCP server and the REST server run on your machines. See What runs where.
Does Meaning Memory send anything to a model provider, or to you?
Section titled “Does Meaning Memory send anything to a model provider, or to you?”Storing one memory and searching never call a language model. Three parts use models, and all of them run where you choose:
- Semantic search on Postgres uses BGE-M3, an open embedding model, through an embedding service you host (
MM_EMBEDDING_SERVICE_URL). - Extracting memories from notes during consolidation calls the provider you configure: OpenAI, Anthropic or a local model through Ollama.
- Scoring a bulk import.
mm_bulk_rememberandmeaning-memory seedscore each memory’s STARE dimensions with a small model that runs in-process: a bundled adapter on an open 1.5B-parameter base model, which is downloaded from Hugging Face the first time it is used. Passauto_stare: falsetomm_bulk_remember,--no-auto-staretoseed, or give the scores yourself to skip it. The model needs the optionalmeaning-memory[stare-bundled]extras. If the model can’t load, the memories are stored without those scores.
The license check is done locally and doesn’t contact us.
Which agent frameworks work with it?
Section titled “Which agent frameworks work with it?”Anything that speaks MCP or HTTP:
- MCP: Claude Code, Cursor, OpenClaw, Hermes and LangGraph have setup pages under Connect your agent. Other MCP clients use the same
mm-mcpserver. - REST: any runtime can use the REST API.
How is this different from a vector database?
Section titled “How is this different from a vector database?”A vector database finds text that looks like the query. Meaning Memory also tracks how much each memory matters, keeps promises on a due list until they are closed, records where every memory came from and every change to it, and compiles a ranked summary sized for the agent’s context window. Semantic search is one part of it. See What is Meaning Memory? and STARE.
Should I start with the file backend or Postgres?
Section titled “Should I start with the file backend or Postgres?”Start with the file backend for one agent: no database, nothing to run. Move to Postgres when several agents need shared memory, semantic search, or scope groups. Export and import moves you from one to the other.
Can several agents share memory without seeing everything?
Section titled “Can several agents share memory without seeing everything?”Yes, on Postgres. Each memory is private to its writer unless it is written into a scope group, and an access policy says which agents belong to which groups. See Scoping and sharing memory.
Can my agent remember what happened in another channel?
Section titled “Can my agent remember what happened in another channel?”Yes. Throughline captures each chat, cron job or background cycle as a short record the agent can find from any other channel.
What happens to my memory if we part ways?
Section titled “What happens to my memory if we part ways?”You keep it. meaning-memory export writes everything to plain JSON Lines files that any tool can read: the memories, their full change history and the links between them. The store stays in your infrastructure either way. See Export and import.
Which features are production-ready?
Section titled “Which features are production-ready?”Significance and time-awareness are the most mature parts. Others are built but switched off until they are proven. Ask the engine: mm_stare_readiness reports each capability’s status from the code itself. See What ships today.