Katami

Developer Stanko released katami, an open-source memory tool that brings Hermes-like memory extraction and recall to Claude Code, Codex, pi, and OpenCode by hooking into each harness to intercept messages, agent responses, and tool calls. Katami spawns a small agent every dozen or so messages to extract memories from the transcript, categorizes and stores them in a local database, and uses fast local CPU AI models to select and inject only the most relevant memories as context, with a mesh-based sync that shares memories across multiple machines. The tool is designed so users do not have to repeat their preferences in every new agent session.

Over the last few months I've built a couple of tools to help me work with our new AI friends. My favorite one so far is katami https://github.com/monorkin/katami Japanese かたみ for "keepsake" . The best way to explain what katami does is with an example: bash $ mkdir hey2 $ cd hey2 $ cargo init $ katami claude Can you set this project up for me? It will be a server with a CLI connected to MySQL, Redis and Elasticsearch I'll set this project up with a hey2-core crate for logic and a hey2 crate for the CLI using usage-rs, configure AGENTS.md with a CLAUDE.md shim, .agents/skills and a docs folder, add a .mise.toml file and set up mbx just like in your other projects. ... Didn't catch that? Claude, without being prompted , in a brand-new project, knew how I want the project structured, what tools and libraries I use, and how I like them configured. The kicker: I didn't tell it to remember how to set up projects, and it's not in my global CLAUDE.md file either. Claude, through katami, learned what I like and remembered it. I built it so that I wouldn't have to repeat myself in every Claude session I started. Katami brings Hermes-like memory extraction and recall to Claude Code, Codex, pi, and OpenCode. You can switch harnesses and machines and keep the same memories. Katami hooks into your harness. It sets up a few temporary hooks which allow it to intercept your messages, the agent's responses, and tool calls. It extracts memories by checking the transcript for things you revealed about yourself. Every dozen or so messages, it spawns a small agent and asks it to comb through those messages and extract memories. The memories it finds are then categorized in a few ways and stored in a local database. bash $ katami memory list id updated kind uses last used title cwfbzcq2 2026-09-02 card 188 2026-10-06 /home/stanko/Work/basecamp/hey2 kfz0scd8 2026-09-02 card 143 2026-09-30 Stanko dje5yqzr 2026-09-08 observation 91 2026-10-06 Uses HEY email CLI hhe3cghc 2026-09-13 observation 63 2026-10-06 Codex review workflow rxtsz6tx 2026-09-24 status 14 2026-09-30 Current state of /home/stanko/Work/monorkin/katami It doesn't blindly stuff memories into your agent's context. Using fast local AI models that run on the CPU, it searches for and selects the most relevant memories for the conversation and injects them as extra context for your agent when it's about to respond. Katami also periodically curates its memories: drops unused ones, combines related ones, and builds fact cards. This is what a memory looks like: bash $ katami memory show kfz0scd8 Stanko card entity: person:Stanko updated: 2026-09-30T08:05:13Z Preferences Communication - Use plain language: avoid vague umbrella terms like "gate" for CI checks or scope validation; use specific plain words for what each thing is - Be explicit about outcomes: say "tests ran and reported failures" not "passing" — lead with the result first, analysis after - Preserve full output for accountability: don't truncate failure lists or logs for readability e.g., avoid tail -150 — there should be no guessing about what happened Git operations - Require explicit permission before force-push; use --force-with-lease when approved Recently we started using agent sheds at work. A shed is a mini-PC that's on 24/7 and hosts an agent that you give tasks to. So I extended katami to work across multiple machines. It uses a mesh, so there's no server to manage. Any extracted memories are shared across all machines. That's it. I love this tool. It makes any old agent feel more human. It allows them to learn and adapt. Every new session feels like another conversation with a digital friend, rather than meeting an intelligence that's just been spawned into existence.