Memory for your coding agent that gets sharper every run
Kimetsu is a single Rust binary that runs next to your agent over MCP. It remembers what matters, learns which memories actually helped, and lets that knowledge compound across sessions. Proactive memory for AI coding agents.
$ npm install -g kimetsu-aiThe parts worth a closer look
Most memory tools are a bucket you read and write. Kimetsu is a cycle: every session feeds the next one, and the knowledge that proves useful is the knowledge that survives.
- 1
Work
The agent solves a task and earns a lesson worth keeping.
- 2
Capture
It lands in the brain as a memory. No model, no cloud, no cost.
- 3
Recall
Next run it comes back by meaning, before the work even starts.
- 4
Cite
The agent marks the memories that actually moved the task.
- 5
Learn
Cited memories rise in rank; stale, unused ones decay and prune.
Most memory waits to be asked. Kimetsu speaks first.
A vector store sits idle until you query it. Kimetsu is proactive: it reads the moment, decides what the agent needs, and puts it on the table before the first turn, so knowledge shows up when it changes the outcome, not after.
you ask → it fetches
Nothing happens until a query arrives. If the agent does not know to ask, the memory never surfaces.
it anticipates → then delivers
The brain surfaces the right context on its own, at session start and ahead of each task, before the agent can repeat a solved mistake.
Session-start digest
Open a new session and the brain has already handed the agent the conventions, gotchas, and decisions that matter here.
Episodic resume
It picks up where you left off: what you were doing last time and the state you left it in, without you re-explaining.
Context before the ask
Before a non-trivial task, the relevant lesson is surfaced automatically, so the agent avoids the mistake instead of repeating it.
Remembers what matters
Project conventions, failure patterns, the exact command that regenerates your schema. Captured once, retrieved by meaning.
Learns what helps
Memories the model cites before it solves a problem get promoted. Stale advice and silent passengers decay and get pruned.
Never explores twice
A session-start digest and an episodic resume mean the first turn already knows the repo and what you were doing last time.
Answers, not just injects
kimetsu ask composes a grounded, cited answer from memory using a local model. Zero frontier tokens, works offline.
Model-free retrieval
FTS5, local embeddings, and a local cross-encoder reranker. Nothing in storage or retrieval calls an LLM.
Yours on your machine
One SQLite file per project. No external vector database, no cloud, no telemetry. Back it up with cp.
Export, merge, and swap brains freely
A brain is a file, not a hostage. Pack up what your agent has learned, hand it to a teammate, install one from a URL, or swap whole brains in and out. Onboard a new machine or a new hire with one import.
Export
One command turns your brain into a shareable pack: gzip-compressed and security-scrubbed, so credentials and PII never leave your machine.
Merge
Import a teammate’s pack additively. It dedups against what you already know, and re-importing is always safe.
Swap
Replace your current memories with a pack’s, reversibly. Old memories are superseded, never deleted, so you can always swap back.