Not a chatbot. An agent that runs on hardware you own.
It searches the web, runs tools, curates its own memory, and shapes its persona around how you talk to it. Self hosted, private by default, fully yours.
Private by default. No cloud dependency. The model, homeserver, search engine, and vector store all run on a machine you control.
Model, Matrix homeserver, search engine, and vector store all run on your hardware. Your data never leaves the box.
Every conversation is saved and embedded for on-device search. Recall a chat from weeks ago by meaning - or pull up a whole day or week by date.
A background deriver distills conversations into confidence-scored conclusions and auto-curated profile cards. It also learns about the people, projects, and places you mention and tracks dated events, surfacing upcoming ones proactively. Only relevant memories are recalled, contradictions resolve to the stronger fact, and weak ones decay over time.
41 built-in tools on an iterative reasoning loop. Web search, files, git, downloads, email with a reactive inbox, SSH remote admin with file transfer, self-authored skills, background workers. Drop a Python file in to add your own.
Run local models through Ollama, reach 200+ cloud models via OpenRouter, or point at any OpenAI-compatible endpoint. Discover available models from the panel and swap providers with one config change. Give any model a price and /costs tracks token-based spend for it.
Spin up specialist sub-agents with their own identity, memory, and Matrix presence - from chat or the control panel. They can consult each other on their own.
A local panel at localhost:8800. Edit config, test connections, manage secrets, and curate memory without touching JSON.
Non-root container, read-only root filesystem, all capabilities dropped, SSRF guards, and human-in-the-loop on destructive ops. Web search and fetched page content is screened for prompt injection by a local classifier before it reaches the model.
A built-in SearXNG instance for searches that do not phone home. Fetch, read, and summarize any URL, fully self hosted.
Ask how Memtrix works and it answers from its own docs. The search_docs and ask_docs tools research the bundled documentation, with sources.
Acts as a sysadmin over SSH: its own key, a persistent shell where state persists between commands, trust-on-first-use host keys, and confirmation for destructive commands.
Memtrix learns from multi-step tasks and captures reusable workflows as skills. On future requests it sees a catalog of all its skills and loads the relevant ones on demand.
Flip on a model's vision toggle and Memtrix sees the images you send it — describing, reading, and reasoning over pictures directly. Works across Ollama, OpenRouter, and any OpenAI-compatible endpoint.
Clone, configure, launch. The setup script and onboarding wizard handle the rest.
then open Element, connect to localhost:6167, and invite @memtrix:memtrix.local to a room. Prefer a UI? Open localhost:8800.
Every component runs locally in Docker on a private network. Only the control panel is published, bound to localhost.
The Python agent. Orchestrates model calls, tool execution, memory, sessions, and sub-agent management.
FastAPI backend plus a React control panel for config, secrets, connection tests, and memory administration.
Standalone vector store shared by the agent and the panel. Powers semantic recall and reasoning memory.
Lightweight Matrix homeserver, local only and no federation. Swappable for any external homeserver.
Privacy-respecting metasearch engine for web access. Self hosted, no tracking, no leaks.
Local inference through Ollama, 200+ cloud models via OpenRouter, or any OpenAI-compatible endpoint (llama.cpp, vLLM, LM Studio, OpenAI) when you need the extra reach.
Every tool is a drop-in Python file. Add one and the agent picks it up on the next restart.
Searchable conversation history captures everything verbatim. A reasoning layer distills durable knowledge into auto-curated cards.
Every conversation is saved as a raw transcript, chunked, and embedded on-device with nomic-embed-text-v1.5 so the agent can search its whole history — by meaning, or by a specific day or date range. No journaling by hand — a persistent hash cache means restarts re-embed only new chunks.
A background deriver turns raw chat into confidence-scored conclusions about you, then re-curates a compact profile card. Only relevant memories are injected, a nightly pass resolves contradictions and lets weak facts decay, and the agent reads the card every turn but cannot overwrite it by hand. The persona genuinely evolves.
The web control panel lets you edit config, test connections, rotate secrets, and browse or edit memory, then apply changes with a safe restart.
Private, persistent, and entirely under your control. No subscriptions, no telemetry, no data leaving your machine.