1.77x
One Kimetsu brain
Better answers than no memory (+77%; Qwen 7B on LoCoMo, 40.3 vs 22.8).
Keep project decisions, conventions, and fixes between sessions. Kimetsu stores them locally and retrieves the context your agent needs.
Kimetsu Network
The same project lessons follow you from machine to machine, in Grok Bot, Codex, Claude Code, Cursor, and the Pi package. Each machine keeps its own local SQLite brain and exchanges a signed, scrubbed repo pack only when you ask. Nothing is listening until you run share or request, and there is no server to stand up.
How Kimetsu Network moves a lesson from Agent A to Agent B
5. B answers, with a citation. Agent B answers using the imported lesson and cites the memory that helped (kimetsu_brain_cite).
Reduced motion is on, so steps fade in place instead of moving. Use Play or the step buttons to move through the story.
kimetsu network invite --name laptopOn machine A, in the repo. Copy the code (it starts with knet1.). The default name is friends. Kimetsu is not listening yet.
kimetsu network join <code>On machine B. Membership is saved. Still nothing is listening.
kimetsu network shareLeave share running on A. On B: kimetsu network request. If discovery misses A, pass the peer: line share printed.
1.77x
One Kimetsu brain
Better answers than no memory (+77%; Qwen 7B on LoCoMo, 40.3 vs 22.8).
22% to 65%
As a network
Each agent answers better with merged memory than with only its own (Qwen 7B and 0.5B, LoCoMo and LongMemEval).
about 99%
Over QUIC between machines
Of the gain kept (7B LoCoMo 40.1 vs 40.3).
Token F1 on LoCoMo (answerable questions) and LongMemEval (proxy F1, not the official judge). QUIC measured on one machine over loopback. Tables and caveats
Install commands in Get started
v2.11.0 prebuilt and npm binaries include kimetsu network. cargo install kimetsu-cli still needs --features network. Relay is off, and router port mapping stays off unless you pass --map-ports. This is not Kimetsu Remote, and Claude Desktop has no installer. The Pi package can request a pack; it does not start share.
73.3%
Accuracy · 400 probes · 20 conversations
Both runs use the same test set.
BEAM test settings →89.4%
Accuracy · 1,540 questions
Questions about facts, events, and relationships across long conversations.
LoCoMo results and comparison →83.0%
Accuracy · 200-question stratified slice
Sampled from the 500-question set, covering recall, changed facts, and reasoning across sessions.
LongMemEval test settings →Historical results from the linked runs. An LLM answers and grades the questions; the memory pipeline uses local search, embeddings, and reranking, with no LLM calls.
Select your agent and run these commands from your project directory. Available for Linux, macOS, and Windows.
All installation options →npm install -g kimetsu-ai
kimetsu setup --host claude-codeRequires npm. Installs the lean build; local semantic models are optional.