ZAM (Zusammen) 🤝
Do real work with your AI — and keep the knowledge instead of losing it.
ZAM is Bavarian for "together".
ZAM turns everyday work with your AI agent into active-recall practice, so growing
automation doesn't mean growing dependence. You get the task done and you get
sharper — the two stop being a trade-off.
Don't just automate. Elevate.
🌐 zam-os.org — the project website, in 7 languages.
Who ZAM is for
Anyone who works with an AI agent and doesn't want to get rusty as it takes on more —
whether you're learning the field (say, a Fachinformatiker apprentice building durable
competence), sharpening your craft on the job, or simply keen to keep growing while you
automate. If you're pairing with Claude, Codex, Copilot & co. anyway, ZAM makes that time
compound into lasting skill.
What ZAM does today
- Rides along in your AI agent. As you work a real task, ZAM breaks it into small
knowledge concepts, notices which ones you're due to revisit, and weaves them into
the session.
- Watches you work. Do a step well on your own and ZAM quietly marks it learned — no
interruption. When no real task can show what you know, it asks a focused recall
question. Both are active recall.
- Remembers what you're forgetting. Every concept is scheduled with FSRS-6 spaced
repetition, including short learning and relearning steps, over a prerequisite graph.
- Stays on your machine. One local SQLite database (
~/.zam/zam.db), shared by the
agent and the Desktop Studio. Review works offline; local LLMs (Ollama, FastFlowLM)
are supported.
Two places to use ZAM
Your agent app is the main workbench. ZAM Desktop Studio is for setup, content,
and focused review. They share the same local database, so progress in one shows up in
the other.
1. In your AI agent — the workbench
This is where the real learning happens: turning actual tasks into practice, observing
your work, and guiding you step by step. ZAM connects to the agent apps you already use:
| Agent | Connect with |
|---|
| Claude (Code / desktop app) | zam agent connect claude-code |
| Codex | zam agent connect codex |
| ZCode | zam agent connect zcode |
| Antigravity | zam agent connect antigravity |
| OpenCode | zam agent connect opencode |
| GitHub Copilot (CLI / app) | zam agent connect copilot |
| Goose | zam agent connect goose |
| Grok Build | zam agent connect grok |
| Hermes | zam agent connect hermes |
One command writes the MCP config (your agent may ask you to approve the server). For
GitHub Copilot, it also installs user-scoped Studio, Recall, Graph, and Settings canvases;
restart Copilot or start a new session after connecting. Then just type /zam — or
say "let's do this together with ZAM" — and work normally.
Portable Agent Plugin
The repository and published npm package also follow the vendor-neutral
Agent Plugins v1.0.0 format. Compatible clients can load
the package root to discover the ZAM skill and stdio MCP server together. A source
checkout needs npm ci && npm run build first; published packages already contain the
runtime. See ZAM Agent Plugin for the layout and validation.
2. ZAM Desktop Studio — setup, content & atlas
A native app (zam ui) for the things a chat window isn't good at:
- Guided setup — the first start connects your AI model, agent, and workspace one
page at a time; everything stays editable in Settings later.
- Easier configuration — pick your language and local AI model in a settings panel,
not a config file.
- Import your own material — paste notes, point ZAM at a source, or walk a structured
curriculum; a guided wizard turns any of them into review cards.
- Edit your content — a real editor for concepts, questions, and prerequisites.
- See your knowledge atlas — what you know as a map of islands: the colour says the
subject, the texture how well you remember it, and red threads show what builds on what.
Zoom from a subject down to one concept, and review a whole region from there.
- Review — run focused active-recall rounds right in the app.
zam ui # launch the Studio
zam ui --build # one-time: build a native installer (needs Rust)
Review works in both places. Observation and guided task-work happen inside your agent.
Quickstart
1. Get ZAM. One line installs the desktop app and the zam CLI:
# macOS · Linux
curl -fsSL https://zam-os.org/install.sh | sh
# Windows · PowerShell
irm https://zam-os.org/install.ps1 | iex
Or grab an installer from Releases, or build
from source (git clone → npm install && npm run build).
2. Open ZAM. The first start walks you through setup, one page at a time: who you're
learning as, your AI model (cloud or local), your agent, your workspace, and your first
learning content — from a curriculum, your own sources, a project, or a goal you name.
Every page can be skipped and finished later; the dashboard keeps the remaining steps as
a checklist. No terminal required.
3. Learn while you work. Open your agent, start a real task, and type /zam. It
checks what's due, plans the concepts behind the task, hands you the work, watches how it
goes, and updates your schedule. Prefer a gentler start? Import material and do a review
round right in the Studio.
Multi-device (a server database and mobile pairing) is an optional later upgrade in
Settings — the first run stays fully local.
Prefer the terminal?
The same setup runs as commands — zam init is the guided wizard, CLI-style:
zam init # workspace · AI model · database · /zam skill
zam agent connect claude-code # or codex · zcode · grok · antigravity · opencode · copilot · goose · hermes
How it works
- Token — one atomic concept worth remembering, tagged with a Bloom level (1 remember → 5 create).
- Card — your personal spaced-repetition state for a token (FSRS-6).
- Prerequisites — a graph of what must be understood first; ZAM won't quiz a concept
whose foundations you've just forgotten.
- Sessions — every work/learning episode is logged, so ratings come from real evidence.
The learning engine is an AI-agnostic kernel with zero LLM dependencies; the agent
layer just drives it. See Kernel Architecture.
Documentation
License
Apache 2.0 — see LICENSE.
