Claude Code, Cursor and any other MCP app can store what they learn together with the sources behind it, find it again in plain English, and remove a bad source and everything built on it. No code to write and no server to run: the app starts CheeseDB itself.
go install github.com/govinargo-labs/cheesedb/cmd/cheesedb@latest
cheesedb agent create alice
claude mcp add cheesedb -- cheesedb mcp alice
Other apps (Cursor, Claude Desktop): add {"mcpServers":{"cheesedb":{"command":"cheesedb","args":["mcp","alice"]}}} to the app’s MCP settings. Then tell the model to keep its long-term memory in CheeseDB.
Add an embedding model to ~/.cheesedb/llm.conf and restart the app. The model gets a sixth tool, search, and every memory it saves becomes searchable. Without it, the five other tools work exactly the same.
embed_base_url = http://localhost:11434/v1
embed_model = nomic-embed-text
add_lotRegister a new outside source: a message, a web page, a tool result.rememberStore something learned, with the ids it came from.searchFind memories by meaning, best first. Never returns a recalled one. (With an embedding model.)recallRemove a source and everything built on it, and erase the text.getRead one lot or memory by id.list_memoriesList what is still alive, newest first.The agent is fixed when the server starts, so a prompt-injected model cannot reach another agent’s memory through a parameter.
The text and search vector of a recalled source and of what depended on it are removed from disk, and recalled text is never shown back to the model.
It runs on your machine over standard input and output. No network port, no account, no dependencies. With a local embedding model, nothing leaves your machine.
Know the limits. The model declares the sources and legs itself, so a recall is only as complete as what it wrote. A model that can also run shell commands can skip the tools and read every agent’s files. See the limits and the threat model.
You talk to your app as usual. The model decides which tools to call; here is what happens underneath.
The model calls add_lot twice (your message, the doctor’s note), then remember “the user is allergic to peanuts”, learned from either one.
The model calls search with your question, gets the closest memories that are still valid, and answers from them.
The model recalls the page by the id add_lot gave it when it saved the page, after you confirm: the page and every memory that depended only on it are erased. The first recall only shows what would go; if your app can show forms, cheesedb asks you itself, and the model cannot answer for you. For a memory you describe in words, it uses search first and shows you what it found.
The allergy is still there: your own message and the doctor’s note still back it. The wrong conclusion about the satay is not.