tiroir

Configurable, scoped memory for AI agents. Tiroir is French for drawer.

Give any AI agent a memory organised the way its user actually works. A developer thinks in repositories, an accountant in client files, a consultant in engagements. With tiroir, the organisation of memory is configuration: you declare your drawers, and tiroir handles storage, indexing, retrieval, context injection and sync.

Documentation: ancelin-code.github.io/tiroir

Status: 0.1.0. One server for a team, run with Docker Compose behind Caddy, administered from the command line, and connected to Claude Desktop with tools to search, read and log memory. More surfaces follow, milestone by milestone.

Principles

• Plain files are the truth. Memory is readable Markdown in a Git repository. Every index is derived and rebuildable.
• Your drawers, your schema. Drawer types, layout, metadata, templates and injection rules are configuration, not code.
• Isolation over recall. When a drawer is active, retrieval stays inside it. Isolation between drawers is a tested guarantee.
• Abstain rather than mislead. Injecting nothing beats injecting a plausible but wrong memory.
• You choose where your memory lives. It syncs to a Git repository you pick, self-hosted or hosted. Nothing else leaves your machine: no telemetry, no third-party APIs, and embeddings are computed locally.
• Git without Git. tiroir syncs your memory across machines and team members through Git, and you never have to touch Git.
• Humans own the synthesis. Logs are append-only, and summaries change only after you confirm.

Quickstart

On a VPS with Docker Compose and a domain name pointing to it:

git clone https://github.com/ancelin-code/tiroir.git && cd tiroir
cp .env.example .env    # set TIROIR_DOMAIN and TIROIR_ACME_EMAIL
docker compose pull && docker compose up -d
docker compose exec tiroir tiroir admin init

Then follow the documentation:

• Install tiroir on a VPS: the stack, the first administrator, a member and a store.
• Connect Claude Desktop: add the server as a custom connector and sign in.
• CLI reference: the administration commands.

Planned surfaces

• A member CLI and the Claude Code plugin with automatic context injection.
• A guided setup skill that interviews you and designs your schema.
• Presets beyond accounting firms, starting with developers.

Contributing

mise run check runs every check, mise run stack builds the image from the working tree and starts the stack, and mise run docs:serve previews the documentation site.

Project documents

• Product brief: vision, principles and constraints.
• Requirements: what v1 must do, checkably.
• Prior art review: what we learned from existing agent memory projects.
• Skeleton exit check: the checklist run against a real server at the end of the first milestone.

Licence

Apache-2.0.