What it does
GoodMemory sits between an application or coding agent and the model it calls. It writes selected facts, retrieves the relevant ones, injects them into the next turn, records what happened, and lets users correct or delete entries. The project states that it is not an LLM, agent framework, vector database or generic RAG system. There are three ways in. Developers can embed the package in a Node or Bun service and use calls such as remember, recall, buildContext, feedback and forget, with an adapter for the Vercel AI SDK. Codex and Claude Code users can install a global CLI, run a setup command, and get hooks that inject context at session start and on each prompt, plus a read-only MCP server. Backends can instead run an HTTP bridge with a Python client, or use a hosted instance. Storage defaults to SQLite, with Postgres and embedding providers available as options. Saving memories from coding sessions is opt-in, with modes ranging from off to selective, an audit ledger, and a command that undoes a given write. The README also publishes benchmark results along with the method behind each number.
Capabilities
- Memory API with remember, recall, buildContext, feedback, forget and export calls
- Setup command that wires memory hooks into Codex and Claude Code
- Read-only MCP server usable from other MCP clients, with an opt-in write tool
- Opt-in writeback modes with an audit ledger and per-write undo
- SQLite storage by default, with optional Postgres and embedding providers
- HTTP bridge with a Python client for backend services
When to use it
- Giving Codex or Claude Code persistent context about a project or user
- Adding user memory to a chatbot, copilot or agent without retraining a model
- Running memory as a sidecar service for a Python or FastAPI backend
- Needing audit and deletion controls over what an assistant remembers
