FAQ
Short answers to common questions.
Yes. It is open source under the MIT license, and the default setup costs nothing to run: no API keys, no hosted service.
Not with the defaults. Only if you choose the OpenAI or Voyage embedding provider. See Privacy.
Any MCP client that can start a stdio server: Claude Code, Cursor, VS Code with Copilot, and others. Without MCP, use remimem context and paste the result.
TypeScript, TSX, JavaScript, Python, and Go at full depth (symbols and call graph). Every other text file is searchable as text.
It returns the functions that answer a question, not every line that contains a word, and it ranks by meaning as well as by name. It adds the surrounding code as signatures, keeps decisions across sessions, and fits all of it into a token budget. On the benchmark it served 86% fewer tokens per task than grep and reading whole files.
No. Your MCP client starts the server when it needs it. remimem watch is optional and only useful without a client.
Commit config.json and memory.json, not the index. remimem init adds the index to .gitignore for you.
Roughly 7 KB per chunk: about 121 MB for a 277,000-line repository.
The local provider runs on the CPU. You can pick another local model, or a hosted provider. See Embedding providers.
Not yet; each person indexes locally. Memories are shared through the committed memory.json. A shared, hosted index is on the roadmap.
remimem counts with the o200k_base encoding. Each model tokenizes slightly differently, so treat counts as close estimates.
It was developed as codemem, but that name was already taken on npm by an unrelated package, so it was renamed to remimem before the first release.
Roadmap
- More languages at structural depth: Java, Rust, C#.
- Team mode: a shared index and memory with access control.
- A pull request reviewer that uses project memory and conventions to flag violations.
- Smarter context: reranking with a small cross-encoder, and learning from which results agents open.