Memmy Agent is a local-first personal memory hub that gives several AI agents one shared long-term context. With your permission it reads local conversation histories from tools such as Cursor, Claude Code, and Codex, and organizes preferences, decisions, project context, and lessons from earlier work into searchable memory.
When you work in a connected agent, Memmy supplies only the memories relevant to the current task, so switching between tools does not mean rebuilding context. Connected agents include DeepSeek Harness, OpenClaw, Hermes, Claude Code, Codex, Cursor, OpenCode, and Pi, which receive a Memory Skill and a hook or plugin when you opt in.
Memmy is also an agent in its own right: you can hand it a task, and it draws on stored memory to organize material, merge plans discussed across agents, or turn past decisions into a document. It ships as a desktop app, a CLI and terminal UI with an OpenAI-compatible API, and a standalone memory service CLI for scripts and other agents.
Memory is stored on your device by default. You choose which sources it reads, can stop collection, view and manage individual memories, and decide which AI tools may use it. Agent tasks run on trial credits after sign-up or on your own model API keys. Planned work extends memory sources to browser activity and local documents, and adds agent-to-agent sharing for teams.
Features
- Cross-agent memory: one shared memory used by Claude Code, Codex, Cursor, OpenClaw, Hermes, OpenCode, Pi, and other agents
- History import: scans local AI conversation histories, with permission, and extracts preferences, decisions, and project context
- Task-relevant recall: injects only the memory that matches the current task, not the whole history
- Built-in agent: runs tasks such as merging plans or writing documents from stored context
- Desktop app, CLI, and TUI: several entry points sharing the same memory and agent runtime
- OpenAI-compatible API: serves the agent runtime to other clients
- Memory service CLI: search, add, and fetch memories from scripts and agents, or run the memory service alone
- Local-first control: memory stored on the device by default, with per-source and per-tool access controls
- Bring your own key: switch to your own model API once trial credits run out
