A self-hosted personal AI agent with persistent memory, powered by local inference via Ollama.
MemoryClaw runs entirely on your own hardware. It remembers your conversations, searches the web, manages your email and calendar, transcribes voice input, and can execute multi-step plans autonomously — all without sending your data to third-party services (unless you explicitly enable optional integrations).
- Persistent memory — hybrid vector + keyword search over everything you tell it
- Local inference — runs any Ollama-compatible model on your own GPU/CPU
- Web search — private meta-search via built-in SearXNG instance
- Email and calendar — read/send Gmail, read/create Google Calendar events
- Voice input — speech-to-text via Whisper, processed locally
- Build mode — plans and executes multi-step goals with tool access
- File understanding — upload PDFs, DOCX, XLSX, images, audio, ZIP archives
- Identity system — customisable personality, rules, and context via markdown files
- Scheduled tasks — morning digest, automatic memory distillation, custom cron jobs
- Claude Code integration — optional access to Claude Code CLI for complex coding tasks, with strict privacy boundaries
- Docker with Compose v2
- Ollama installed on the host
- 8 GB+ RAM (more for larger models)
git clone <repo-url> memoryclaw
cd memoryclaw
./install.shSee docs/SETUP.md for a detailed installation walkthrough, docs/CONFIGURATION.md for full configuration reference, and docs/ARCHITECTURE.md for technical details.
| Document | Description |
|---|---|
| Setup | Installation and first boot |
| Configuration | Identity, memory, tools, environment, scheduler |
| Architecture | Technical deep-dive for developers |
| Integrations | Google OAuth, Claude Code, custom tools |
| Troubleshooting | Common issues and fixes |
| Changelog | Release history |
MIT