Share coding agent memory across sessions and devices
If you stop using Claude Code and open OpenAI Codex in the same folder, you do not need to re-explain the architecture or failed approaches. This tool collects the records left by each agent into a single server so the next session can pick up where the last one left off. It changes the structure from memory stored separately for each agent to a shared format.
Memory that continues across agents and devices
The memory of existing coding agents is usually tied to a single computer and a specific agent. You can open a project interrupted on a desktop on a laptop and carry over the same knowledge and unresolved questions. If team members share a single server, one person's agent can search for content learned in another person's session. Personal handoffs are maintained individually, while project knowledge is shared.
A structure where Markdown files are the source
The source of memory is a git-based wiki, stored as plain .md files. The database is merely an index derived from these files and can be recreated at any time. There is no need to manage a separate vector store or have content trapped in binary files. You can search with grep or open and edit directly with Obsidian. Deleted content can be restored via the restore-page command through git history and version chains, reducing the risk of data loss.
A basic mode that works without an LLM
In the default configuration, capture, search, and handoff all work without an LLM API key. Lifecycle hooks automatically record prompts and tool calls, refining them at a personal information boundary before storage. When a session ends, these observations are organized into readable Markdown pages. Memory naturally decays based on usage frequency, with frequently used content retained longer. Similar content is merged into one, and conflicting content is flagged. This entire process occurs without API calls.
Installation and agent connection
It supports Linux, macOS, and Windows (WSL2), and can be installed via Docker or native binaries. Arch Linux offers an AUR package, Fedora offers an RPM, and macOS offers a menu bar app or source build. After installation, connect agents using the ai-memory install-mcp and ai-memory install-hooks commands. The server binds to the loopback address by default to block external access; if exposing it to a LAN, you must add bearer token authentication.
Things to check before use
This tool is distributed as a single binary and has multi-user authentication and audit logs built in by default. LLM-based 'dream' features and vector search are optional and disabled by default. Native Windows support is experimental, so using WSL2 is recommended. In team environments, you should pre-configure server access permissions and audit log policies. The README does not specify the license type, so you must directly verify the license terms before applying it to commercial projects.
