Kage
one learns, all recall
Last verified:
What is Kage?
Kage is a Git-native memory framework for coding agents that helps teams stop re-explaining their codebase and avoid acting on stale knowledge. It stores learnings as versioned files in the repository, so memory is reviewed in pull requests, shared across the team, and tied directly to code.
The core idea is that every lesson an agent learns becomes a verified memory packet. Kage checks citations against the repo before writing them, rejects unsupported claims, and withholds stale memories when referenced files have been deleted or changed. This makes the system behave more like a trusted layer of project knowledge than a generic note store.
Kage is designed to work with agents such as Claude Code and Codex through MCP, with support for other agents via paste-in snippets. It can auto-wire setups, inject verified memory at session start, recall memory when files are read, and warn when a diff invalidates existing team memory before the PR lands.
It is aimed at developers, teams, and agent-heavy workflows where repeated context loss wastes time and tokens. The website emphasizes local-first use, Git-based collaboration, private sync over your own remote, and no account or API key for the open-source core.
Kage pricing
Pricing model: Freemium
The open-source core is listed at $0 forever and includes verified memory, Truth Report, receipts, auto-capture, repair, live viewer, support for 15 agents, and kage sync over your own private git remote. Kage Cloud is listed at $0 early access and adds memory that follows you on every machine behind one private MCP link, while verification stays client-side. A Team plan is coming soon and is described as shared team memory with review gates and hosted collaboration; the site says it is not yet available.
Kage pros
- Git-native memory stored as plain files
- Verified against the codebase before writing
- Rejects hallucinated citations
- Withholds stale memories at recall time
- Catches invalid memory before PR merge
- Versioned in git and reviewed in pull requests
- Shared across the whole team
- Works with Claude Code and Codex
- Supports other MCP-based agents
- Zero-config installation path
- No account required for the core
- No API key required for the core
- Local-first and private by default
- Personal memory syncs across machines
- Token savings receipts per recall
- Live dashboard for memory activity
- Open-source core is free forever
- Can audit existing claude-mem stores read-only
- Grounded memory injection at session start
- Checks memory against repo changes over time
Kage cons
- Best fit is coding-agent workflows
- Depends on a Git workflow
- Requires MCP-compatible setup for best use
- Team cloud offering is still coming soon
- Early access cloud is limited
- Initial setup still requires command-line use
- Memory quality depends on repository citations
- Designed for code repos, not general note-taking
- Personal sync uses a private git remote you own
- Some features are strongest in Claude Code
Frequently asked questions about Kage
What does Kage actually do?
Kage gives coding agents persistent memory that is checked against your repository. It stores learnings as git-tracked files, verifies citations before writing them, and helps agents recall the right project knowledge instead of rediscovering the same context repeatedly.
How does Kage avoid stale or incorrect memory?
Kage validates every memory packet against the repository before it is saved. If a memory cites a file that does not exist, it is rejected, and if the referenced file is later deleted or changed, the stale memory is withheld from recall.
Which agents does Kage support?
The site says Kage auto-wires Claude Code and Codex during install, and it also works with other agents through a paste-in MCP snippet. The website also lists broad support for many MCP clients, including Cursor, Windsurf, Gemini CLI, Cline, Goose, Roo Code, Kilo Code, OpenCode, Aider, Claude Desktop, Copilot, OpenClaw, and Hermes.
Do I need an account or API key to use Kage?
No for the open-source core. The site explicitly says the core runs with no account and no API key, and that verification stays local and private by default.
How is Kage installed?
The website shows a one-command install using npx -y @kage-core/kage-graph-mcp install. It says that command sets up repo memory and the code graph, then you restart your agent.
What does Kage Cloud add?
Kage Cloud is early access and is described as memory that follows you across machines behind one private MCP link. The site says verification still happens client-side, so the cloud does not see your code.
What is the live viewer for?
The viewer is a memory dashboard that shows recent captures, memory by type, a value timeline, a memory-to-code map, recalls per day, recent activity, memory health, and the codebase Kage has mapped. It is meant to make repository knowledge visible and measurable.
Can Kage help teams review changes that break memory?
Yes. The site says kage pr check warns when a change invalidates team memory before the pull request lands, so the team can update or supersede the outdated memory before it causes confusion.
Does Kage work for personal memory too?
Yes. The site says personal memory lives in ~/.kage/memory and syncs over a private git remote you own, with conflicts resolved newest-wins while keeping both versions.
What is Kage’s trust benchmark?
Kage says its trust benchmark is 99/100 on the site, emphasizing that it measures whether memory can be trusted rather than just whether it can be recalled. The demo and benchmark sections show write-time rejection of hallucinated citations, stale-memory withholding, and grounded recall as the main checks.