jot

One memory for all your AI coding agents, kept as plain Markdown files on your machine.

Tell Claude Code you use pnpm, and Cursor already knows it tomorrow. jot is a small local MCP server that every agent you connect reads from and writes to. No cloud, no account, no database: one .md file per memory, versioned in git.

Quickstart (about 60 seconds)

1. Install the CLI (macOS or Linux):

   curl -fsSL https://raw.githubusercontent.com/ArpitK24/jot/main/scripts/install.sh | sh
   

   On Windows, in PowerShell:

   irm https://raw.githubusercontent.com/ArpitK24/jot/main/scripts/install.ps1 | iex
   

   Both download the newest release from GitHub, check its SHA-256 against the release's checksums.txt, and install one binary. Other ways are under Install.

2. Connect your tools:

   jot setup
   

   It finds Claude Code, Cursor, Codex, Antigravity and Claude Desktop, prints the exact lines it will add to each config, and asks once before writing.

3. Restart those tools. That's it. From now on they are told to recall your preferences before a task and to save new ones as you work.

To check that everything is wired up, run jot doctor.

What it looks like

Monday, in Claude Code:

you:    we use pnpm in this repo, never npm
claude: Noted. (calls jot remember)

jot writes ~/.jot/20261109-101502-3a1c.md:

---
id: 20261109-101502-3a1c
kind: preference
tags: [js, tooling]
source: claude-code
created: 2026-11-09T10:15:02Z
updated: 2026-11-09T10:15:02Z
---

Use pnpm, never npm, in JS projects.

Tuesday, in Cursor:

you:    add a lint script
cursor: (calls jot recall "add lint script to package.json")
        [20261109-101502-3a1c] (preference, from global memory, tags: js, tooling)
        Use pnpm, never npm, in JS projects.
cursor: Running pnpm add -D eslint ...

Whether an agent calls recall on its own is up to the agent. jot's server tells every agent to, and jot setup also adds a short block to the instructions file that Claude Code, Codex and Antigravity read every session. Cursor and Claude Desktop rely on the server's instructions alone.

FAQ

Does anything leave my machine? No. The jot CLI has no server, account or telemetry; your agents start it as a local process. The only network call it can make is to Ollama on localhost, if you turn on semantic search. (The Windows desktop app also checks GitHub for updates when it opens.)

Where do the files live? Global memory is in ~/.jot (or $JOT_HOME). Project memory, if you create it with jot init --project, is in <repo>/.jot, and agents working in that repo search both. jot where prints the stores in use. Each memory is a Markdown file you can read, edit or grep.

How do I undo something? Every change is a git commit in the store (when git is installed). jot log shows the history; git -C ~/.jot show <hash> shows an old version, and git -C ~/.jot checkout <hash>^ -- <id>.md brings back a deleted memory. To disconnect jot from your tools, run jot setup --undo; it removes only what it added, and the first time it edits a config it keeps the original next to it as <file>.jot-backup.

Which tools work? jot setup connects Claude Code, Cursor, Codex, Antigravity and Claude Desktop. Any other MCP client works with command jot, args ["serve"] over stdio; see docs/setup.md.

What do agents save? jot asks them to save few, short, self-contained memories: preferences you state, decisions, corrections, and non-obvious facts about a project. Not task progress, code, logs or secrets. Similar memories are caught before a second copy is saved, and jot dupes finds the rest.

How do I see or change what's saved? jot ls, jot search <words>, jot edit <id> and jot rm <id>, the Windows desktop app, or any text editor.

Install

macOS and Linux: the one-line installer above, or download a jot_<version>_<os>_<arch>.tar.gz from Releases. The installer puts jot in ~/.local/bin (/usr/local/bin as root); set JOT_INSTALL_DIR to choose another folder and JOT_VERSION to pin a version.

Windows desktop app: download the newest jot-<version>-windows-x64-setup.exe from a jot desktop release on the Releases page. It installs the app and the jot CLI for your user, adds jot to PATH and connects your AI tools. The build is unsigned, so SmartScreen may ask you to click More info, then Run anyway. From 0.3.1 the app updates itself: it checks GitHub when it opens and installs a new version in one click, after checking its SHA-256. Light, dark or system appearance is in Settings.

Windows CLI only: the PowerShell one-liner above installs jot.exe to %LOCALAPPDATA%\jot\bin and adds it to your user PATH.

With Go 1.23+:

go install github.com/ArpitK24/jot/cmd/jot@latest

It's also listed on the official MCP Registry as io.github.ArpitK24/jot, with one-click .mcpb bundles on every release.

Connecting tools

$ jot setup
  Claude Code     will connect   ~/.claude.json
  Cursor          connected      ~/.cursor/mcp.json
  Codex           will connect   ~/.codex/config.toml
  Antigravity     not installed  ~/.gemini/config/mcp_config.json
  Claude Desktop  not installed  ~/Library/Application Support/Claude/claude_desktop_config.json
  ...
Connect 2 tools? [y/N]

Besides the MCP entry, jot setup adds a short marked block to the tool's own instructions file, which the agent reads at the start of every session: ~/.claude/CLAUDE.md (Claude Code), ~/.codex/AGENTS.md (Codex) and ~/.gemini/GEMINI.md (Antigravity). Undo both with jot setup --undo. To configure a tool by hand, see docs/setup.md.

jot doctor checks the store, each tool's entry and instructions block, that the configured jot exists and answers MCP requests, and semantic search, and prints the exact command that fixes each problem:

$ jot doctor
...
Tools
  ✓ Claude Code      MCP entry in ~/.claude.json
  ✗                  jot block missing or outdated in ~/.claude/CLAUDE.md
      fix: jot setup claude-code
  ✓ Cursor           MCP entry in ~/.cursor/mcp.json
Server
  ✓ jot serve        ~/.local/bin/jot answers initialize and tools/list (jot <version>, 5 tools)

How it works

~/.jot/                      global memory (or $JOT_HOME)
  20261103-134832-fb9e.md
  .git/
<repo>/.jot/                 optional project memory (jot init --project)

A memory file:

---
id: 20261103-134832-fb9e
kind: preference
tags: [js, tooling]
source: claude-code
created: 2026-11-03T13:48:32Z
updated: 2026-11-03T13:48:32Z
---

Use pnpm, never npm.

Recall uses BM25 ranking with boosts for tag matches and recent updates, across project memories and global ones, and says which store each result came from. An agent started inside a repo with a .jot folder gets that repo's memories; clients that report their workspace through MCP roots get them even when they start jot elsewhere.

MCP tools

Tool	What it does
remember	Save one preference, decision, fact or note. If a similar memory exists, it saves nothing and points to that one instead. It refuses empty text, more than 600 characters, and anything that looks like a key or password
recall	Search memory for what's relevant to the task
update	Rewrite a memory that changed
forget	Delete a memory
list	Newest memories, filtered by kind or tags

CLI

jot init [--project]     jot search <query>     jot show <id>     jot log
jot add [-k -t -p] text  jot ls [-k -t]         jot edit <id>     jot where
jot serve                jot rm <id>            jot version
jot setup [--undo]       jot dupes              jot merge <keep> <drop>
jot doctor

Semantic search (optional)

Keyword search is built in and needs nothing. To also find notes that say the same thing in other words, run Ollama and turn semantic search on:

ollama pull all-minilm   # 46 MB
jot embed on             # or: Settings → Search in the desktop app
jot embed status

Search, recall, jot dupes and the remember duplicate guard then use keyword and semantic matches together. If Ollama isn't running, jot falls back to keyword search. Vectors are cached in your user cache folder, never in the store, and nothing leaves your machine. Use --model to pick another Ollama embedding model (embeddinggemma and nomic-embed-text are tuned too) and jot embed off to turn it off.

Duplicates

Agents often save the same thing twice in different words. jot dupes lists likely pairs, and jot merge keeps the newer one, adds any sentence only the other one had, and removes it in one commit. Git still has the original.

$ jot dupes
 66%  20261106-091200-a3f9  global  claude-code  Use pnpm for every JS project...
      20261104-180511-7be0  global  cursor       Use pnpm for every JS project...
      jot merge 20261106-091200-a3f9 20261104-180511-7be0

Ask (chat with your notes)

jot ask answers a question from your own memories. It recalls the notes that match, sends them with the question to a model you choose, and streams the answer. The desktop app has the same thing as an Ask view, with a "Used N notes" list under each reply and a Remember button to save a fact from it.

$ jot ask "which package manager do we use?"
$ jot ask -m llama3.2 -c <chat-id> "and for Python?"   # continue a chat
$ jot chats                                             # list saved chats
$ jot models                                            # what you can pick

Models: Ollama (local, no key), any OpenAI-compatible server (OpenAI, OpenRouter, Groq, LM Studio) and Anthropic. Keys live in <user config dir>/jot/keys.json with owner-only permissions, never in the store; OPENAI_API_KEY and ANTHROPIC_API_KEY also work, and a saved key wins over them. In the desktop app, set them in Settings → Models.

Chats are Markdown files in the global store's chats/ folder, versioned like everything else. They never show up in Notes, duplicate checks, or what agents get from recall, so a chat can't pollute your agents' memory. Only what you choose to Remember becomes a note.

Roadmap

• Merge near-duplicate memories instead of piling them up
• jot setup to connect installed agents in one step
• Desktop app for Windows (macOS and Linux next)
• Expiry and review for stale facts
• Optional local embeddings (Ollama) for semantic recall
• Ask: chat with your notes using Ollama, OpenAI-compatible or Anthropic models
• jot sync through a git remote across machines
• TUI browser

License

MIT