Works with the AI Agents you already use
Bantamkit Memory
Your coding agent remembers your corrections across sessions — local, offline, no account. Free, MIT.
Free
Bantamkit Memory
Example session with this skill installed
From now on in this project: React components are function components, and props are declared with interface, never type. Remember that. Also remember why Playwright kept failing at random today: EADDRINUSE because npm run dev was still holding port 3000, fixed with reuseExistingServer: true.
- Read your context and instructions
- Compiled the bantamkit memory
Saved both to project memory.
What I stored
| Tool | Memory | Type | Result |
|---|---|---|---|
memory_save | react-interfaces-not-types | feedback | saved 'react-interfaces-not-types' |
memory_save | flaky-e2e-is-port-3000-collision | project | saved 'flaky-e2e-is-port-3000-collision' |
When I tried to save the React rule a second time under a new name, bantamkit refused and pointed me at the existing memory:
similar memory 'react-interfaces-not-types' already exists — save under that SAME name to update it, or skip. Do not rename to force a copy.
Next session: recall before acting
You ask: "add a Button component with a variant prop". Before I write any code, I check memory:
memory_recall {"query": "new Button component props"}
[project] [react-interfaces-not-types] (feedback) react component props: function components, typescript interface never type alias
Write React components as functions. Declare props with `interface`, never `type`.
**Why:** the user asked for it on this project.
**How to apply:** new components and any component you touch.
So the component comes out as function Button({ variant }: ButtonProps) with interface ButtonProps, and you don't have to repeat yourself.
A week later, CI goes red:
memory_recall {"query": "playwright tests failing randomly"}
[project] [flaky-e2e-is-port-3000-collision] (project) playwright e2e tests failing randomly EADDRINUSE: dev server already on port 3000
Playwright e2e fails at random with EADDRINUSE when `npm run dev` is still running. Fix: `reuseExistingServer: true` in playwright.config.ts.
Where it lives
.bantamkit/memory/facts/react-interfaces-not-types.md is a plain Markdown file with YAML front matter. You can read it, edit it, commit it, or delete it.
Every tool output above is copied verbatim from a real run of bantamkit-mcp 0.37.1.
Connects securely to your tools. The creator never sees your data.
About this skill
Every new session, your coding agent starts from zero. It asks for the same preferences again. It repeats a fix you already found. It forgets the decision you made yesterday.
bantamkit memory fixes that. It gives your agent a small, durable memory that lives in plain Markdown files in your repo. Your agent checks that memory before it acts, and saves to it whenever you correct it.
What you get
- Recall before acting. Before a task that looks like past work, the agent searches memory and gets back the matching facts in full, so it starts from what you already know.
- Saves the things that matter. Corrections and preferences (
feedback), ongoing decisions (project), facts about you (user), and links to dashboards or tickets (reference). - No duplicates, no bloat. A save that looks like an existing memory is refused, and the reply names the existing one so the agent updates it instead of piling up copies.
memory_compactandmemory_dreammerge and archive old facts to stay inside a token budget. - Plain files you own. Every memory is one readable
.mdfile in.bantamkit/memory/in your project, with personal facts in~/.bantamkit/memory/. You can commit them, review them in a PR, or delete them. - Private by design. Runs locally over stdio. No account, no API key, and no network calls once it is installed.
- Optional auto-recall in Claude Code.
--install-hooksmakes matching memories show up on every prompt, without the agent having to ask.
Setup (one minute)
npx -y bantamkit-mcp@0.37.1 --install claude # or: cursor | copilot | claude-desktop
Restart your host and the memory tools appear. Node.js 20+ is all you need. A Python build is on PyPI too, as pip install "bantamkit[mcp]".
New in 0.37.1
- Safer skill writes. Temp files are created exclusively, so a symlink planted at the temp name is refused instead of followed (#250).
- Skill fixes. Approving a stored draft re-checks it against the current catalogue, undo after an automatic improvement restores the right version, and a long-running server now finds skills installed after it started.
- Windows. Skill state is replaced atomically. Thanks to contributor @Aschenbath.
New in 0.37.0
- Skill audit. The
skill_auditprompt checks the skills your agent has installed: two skills sharing a trigger phrase, a skill catalogue over its size budget, broken frontmatter, and skills never called in the last 30 days. For a plugin nobody uses, it proposes asettings.jsonchange and never applies it for you. A daily run reports any new finding once, in one short line at session start. - Skills learned from experience (opt-in). bantamkit notices work you repeat successfully and proposes it as an Agent Skill. You review proposals with the
skill_proposeprompt. Automatic install stays off unless you setBANTAMKIT_SKILL_AUTO=on.
Works with
Claude Code, Claude Desktop, Cursor and VS Code Copilot. It also works with any MCP host that runs a stdio server.
Open source
MIT licensed, with source at github.com/Ink01101011/bantamkit. The same tools ship as Node (npm) and Python (PyPI) implementations, and a conformance suite of 19,000+ cases checks that both answer the same way on the same memory store.
Known limitations
- Recall ranks by words (BM25F over name, description and keywords), not by meaning. Write descriptions in the words you will search with later. The skill teaches your agent to do that.
- You need to restart your host once after install.
- It never stores secrets, and never stores things already in your repo or git history.
Requires manual host restart after installation. Does not save data already in git. Avoids secrets. Best for bantamkit v0.36.0.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
1 use
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 18 days ago
- Free with an Agensi account
Requires