What it does
RSI Boot keeps project memory inside the repository, in a `.rsi/` directory, rather than in a chat history or a single tool's account. It stores conventions, prohibitions, repair notes and harness slots. New memories are distilled into pending entries, and only entries a person approves are kept for later recall. Negative feedback goes to a separate failure store and can only change writable slots such as prompts, skills, recall strategy, intent rules and knowledge. The publisher says proposals pass a static gate and human confirmation before promotion, are observed for 7 days, and are rolled back if adoption drops.
The publisher positions it for teams or individuals who use several AI coding tools on the same repository, such as Cursor, Claude Code, Codex, Qoder or DeepSeek Harness, and want shared conventions instead of separate notes in each tool. Memory follows the repository, so a copied or moved repository carries its `.rsi/` folder with it. Approved conventions are also written to a managed block in the root `AGENTS.md` and to `.cursor/rules`. The publisher reports Hit@5 of 93.0% on its own 315-question EngMem-Bench, against 87.3% for a vanilla RAG baseline, and notes that these are retrieval hits rather than model answer scores.
It runs as an MCP server (`rsi serve`) connected to the coding tools. The main path needs no API key and makes no model calls, according to the publisher. By default, learning adds `.rsi/` to `.gitignore`, so logs and caches stay local. To share approved memory, a team must commit `.rsi/memory` and adjust ignore rules, because Git cannot add files from a directory it ignores as a whole.
Limitations and requirements
The publisher's benchmark results are its own measurements and are not yet placed alongside other systems' answer scores; the self-improvement loop has not been measured for whether the next round performs better. Sharing project memory requires committing `.rsi/memory` to the repository and changing ignore rules, since `.rsi/` is ignored by default. Shared memory is limited to the repository; chat history is not included.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
RSI Boot
Repo-bound memory and self-improving harness for AI coding agents — zero API keys.
v0.1.9As shown on registry.modelcontextprotocol.io when we last read it