Skip to main content
Xynthis is an AI brain that lives on your machine and never stops learning you. Where most AI tools are frozen weights in someone else’s cloud that forget you when the session ends, Xynthis stores memory as typed binary records that compound across sessions and trains its own models from them, locally. It is built for people who work in their tools all day and want an assistant whose usefulness grows with every conversation instead of resetting to zero.

What it does

Memory that compounds. Everything you tell it lands as typed binary records in ~/.xynthis/bmc/ and survives restarts; a terminal conversation, a note from the app, and a fact stored via MCP all share one memory. Recall runs a layered retriever in milliseconds, and xynthis ask answers factoid questions straight from memory with zero LLM tokens. Works with your choice of LLM. Fourteen provider presets ship built in, from Anthropic and OpenAI to Ollama, llama.cpp, and the local xynthis-llm server; switch with xynthis models set or the in-app picker. The memory is provider-independent, so you change models without losing anything. An agent, not just a chatbot. The agent loop carries roughly 70 tools: shell, file edits, recall, web search, browser control, and on macOS full computer use that clicks, types, and sees the screen, gated behind the permissions you grant. It pauses to ask you when unsure and delegates focused subtasks to fresh subagents. It gets better every day. A rerank head trains on your real recall traffic, so retrieval sharpens the more you use it. Dream cycles consolidate memory and extract patterns in the background, all on your own hardware. An on-device model that evolves. The bundled local model server exposes an OpenAI-compatible endpoint and fine-tunes itself from brain data, promoting a new checkpoint only when evals pass. Your usage becomes your model, with no training data leaving the disk. Private by architecture. There is no server-side copy of your memory to breach, because there is no server; the brain, the store, and your keys sit in ~/.xynthis/ on your disk. See how the pieces fit together.

What it is not

Xynthis is not a cloud service holding your data. The brain, the memory, the trained rerank head, and your API keys all live in ~/.xynthis/ on your own disk. Nothing is transmitted anywhere unless you point the client at a remote LLM provider. When one is active, each request carries your message plus the query-relevant memory the agent injects before the turn: recalled fact snippets, the working-memory ledger, skill-pattern hints. It never sends the memory store itself. Lite-mode providers skip the injection entirely. API keys sit in ~/.xynthis/auth.json at mode 0600.
The free tier recalls the last 7 days of memory; Pro and Team recall unlimited history. Everything is still stored locally either way. The window applies to recall, not to storage.

Platforms

PlatformWhat you get
macOS (Apple Silicon + Intel)Full experience: menu-bar app, CLI, brain daemon, MCP bridge, computer use
Linux (x86_64 + arm64)CLI + brain daemon + MCP bridge, auto-started via a systemd user unit
Windows (x86_64)CLI + brain daemon + MCP bridge over a named pipe. No app, no auto-start on reboot yet

Quickstart

Install, verify, and have your first conversation in about two minutes.

Why Xynthis exists

Frozen weights forget you. The case for a brain that compounds.

Everything it does

The full feature map across app, CLI, daemon, brain, and MCP.