What is Meaning Memory?
Meaning Memory is a memory layer for AI agents. An agent connects to it as an MCP server, or over REST, and gets tools to store what it learns, find it again later, and keep track of what it has promised to do. You run the engine in your own infrastructure, and the memory stays when the model changes.
The problem it solves
Section titled “The problem it solves”Most agents now remember something between sessions: saved facts, a notes file, a summary of past chats. As they take on real work, that memory runs into the same limits:
- It can’t tell what matters. A passing remark and a standing decision carry the same weight, and outdated facts linger until someone deletes them.
- It doesn’t keep promises. “I’ll send it Friday” is a line in a transcript, not something with a due date that stays open until it is done.
- It can’t say where anything came from. When a memory is wrong, there is no record of who wrote it, when, or what it replaced.
- It stays with one agent, one tool or one vendor. Switch models, add a second agent or change frameworks, and the memory doesn’t come along.
Meaning Memory keeps memory outside the model, in your infrastructure. It decides what matters, keeps commitments until they are closed, records where every memory came from, and lets your agents share it on your terms.
What it gives an agent
Section titled “What it gives an agent”There are four kinds of memory:
- Semantic: facts and knowledge.
- Episodic: what happened, and when.
- Procedural: how to do things. Procedures are always kept in the agent’s compiled memory.
- Commitment: what the agent owes. A commitment stays on the agent’s due list until it is closed, however old it is.
Significance. Every memory carries a significance score that people and agents can raise, lower or pin with feedback. Search weighs it together with age, so an old memory has to matter more to rank as high as a new one. When an agent’s compiled memory has to fit a budget, the least significant memories give way first. Five dimensions describe what makes a memory matter; see STARE.
Commitments that are kept. Promises have due dates, recurring duties have schedules, and closing one is an explicit, recorded act. See Commitments.
Provenance. Every memory records who wrote it and when, and every later change records who made it and why. Any memory’s history can be traced. See Provenance.
A compiled memory file. Consolidation turns the store into MEMORY.md, a ranked summary sized for the agent’s context window, rebuilt on a schedule. See Consolidation.
What runs where
Section titled “What runs where”| Piece | Where it runs |
|---|---|
| The engine, the MCP server and the REST server | Your infrastructure, installed from a Python wheel |
| The memory store | A directory on disk (file backend) or your PostgreSQL database |
| The model your agent uses | Your choice. Meaning Memory does not call it to store or search. |
| Consolidation’s fact extraction (optional) | A model you choose: OpenAI, Anthropic or a local model through Ollama |
Nothing about your agents’ memory leaves your infrastructure unless you send it. You can export everything to plain JSON Lines at any time.
Two backends
Section titled “Two backends”File backend. One directory per agent, no database. Keyword search. The fastest way to start, and a good fit for a single agent or an air-gapped machine.
PostgreSQL. Several agents on one store, keyword and semantic search, and access policies that control which agent sees which memories. The backend for a fleet.
You can move from one to the other with export and import.
Where it fits
Section titled “Where it fits”Meaning Memory sits beside your agent framework, not inside it. Claude Code, Cursor, OpenClaw, Hermes and LangGraph agents connect over MCP with a few lines of configuration; anything else can use REST. See Connect your agent.
For editions and access, see meaningmemory.ai.
Ready to try it? The Quickstart goes from install to an agent that keeps a promise across sessions.