krimto

Krimto is an open-source team memory layer for AI coding agents, with the core goal of returning agents' memory to users' own control. It does not store memory in a vendor's database, but instead keeps it as pure Markdown files in the user's own git repository, so every memory can be read, reviewed in a Pull Request, and has a complete git audit log.

Main capabilities

• Memory is stored as Markdown files in a git repository: each fact is an ordinary Markdown file, located in a user-controlled git repository, can be edited with any editor, its audit trail can be viewed via git log, and it can be reviewed in a PR.
• User -> team -> organization hierarchical governance: knowledge is scoped to an individual, a team, or the entire company, and the most specific scope takes priority during recall. This is a mechanism enforced by the server.
• Cross-vendor, cross-editor synchronization: a single MCP server can synchronize the same memory across multiple editors such as Claude Code, Cursor, Codex, Gemini CLI, and across different machines.
• Three-layer architecture: the storage layer is Markdown files in a git repository (the source of truth, with git as the audit log); the index layer is a hybrid SQLite + sqlite-vec index (keyword + vector) for fast retrieval, and applies scope priority during ranking; the access layer is an API server that enforces who can read and write each scope (user/team/organization).
• Terminal and browser viewing: notes can be viewed in the terminal via krimto notes, or a browser dashboard can be opened via krimto ui. Data is stored in the ~/.krimto directory, independent of any specific project.

How to use

• Quick start (personal use, no account required): setup is completed with a single npx command. The setup wizard detects the user's editor, completes the wiring, and enables automatic memory. After that, telling the agent "remember X" in any chat will save a persistent, attributable fact; when asking a question in a new chat, a different editor, or another machine, the correct answer can be recalled.
• Connecting agents: krimto init automatically wires supported editors. For any MCP client, you can also manually point to Krimto's stdio interface, or use the configuration file form (Cursor, Codex, Gemini CLI, etc. use the same structure)). By default, the agent only uses Krimto when the user requests it; running krimto init once in a project will place a persistent rule that lets the agent use Krimto on its own.
• Install as a Claude Code plugin: If you prefer Claude Code's plugin system, you can add Krimto's marketplace and install it directly. This packages the MCP server along with Krimto's skills, the /krimto-status command, and memory hooks, without needing to run krimto init separately.
• Team sharing: When you are ready to share memory with teammates, there is a corresponding onboarding process that covers the admin email, organization/team name, an optional shared git remote repository, and teammate invitations, and prints a join command for each teammate. Teammates can connect to a shared server, or each run their own Krimto and sync through a shared git remote repository. Personal notes are stored together with team notes and synced as a whole. You can use krimto team disband at any time to return to personal mode, and notes will be retained.

Use cases

• Individual developers want the memory of an AI coding agent to follow them across editors and machines, with data fully under their own control.
• Teams want to share a "team brain" while keeping personal notes private, with permissions governed by user, team, and organization levels.
• Development workflows that need audit trails and reviewable memory content.

Limitations and runtime requirements

• The runtime requires Node 20+.
• The current version v0.2 includes the memory core, team features, a web dashboard, and a cross-vendor MCP server. Future plans: v0.3 will add OAuth login and a Pull Request approval process, and v1.0 will launch hosted Krimto Cloud.
• It is licensed under Apache-2.0, and the same codebase can be self-hosted by individual developers or enterprises, with no tier restrictions.
• More details about specific charges, pricing, platform support scope, etc. are not yet provided in the materials.