Memory
Xynthis stores what you tell it as typed binary records on your own disk, and it stores selectively: a write-time filter rejects noise, every fact carries a trust score, and newer facts supersede stale ones instead of piling up. Background dream cycles consolidate and reflect on what is stored, so memory quality improves with use. The agent can also ingest images, PDFs, and files into memory through the same filter pipeline, andxynthis ingest pulls in iMessage history, open Reminders, and upcoming Calendar events. See how memory works and the working with memory guide.
Recall
Retrieval runs a layered retriever and returns in milliseconds.xynthis ask answers factoid questions straight from memory with zero tokens and no LLM call, warning you when confidence is low. You can cryptographically confirm a recalled fact with your local Ed25519 identity and later print its offline-verifiable audit trail, Merkle inclusion proof included. See truthful memory and recall in the memory guide.
Agent tools
The terminal client runs an agent loop with roughly 70 tools: shell, file read/write/patch, fuzzy file search, recall and remember, and resumable questions back to you mid-turn. The agent reaches the live web too: search runs through DuckDuckGo with no API key, fetch returns stripped page text, and an extraction tool pulls structured JSON out of a page. It reads PDFs via PyMuPDF, runs Python snippets with a 30-second timeout, inspects your clipboard, process list, and network state, sends email through Mail.app or a Gmail compose URL with a send-once guard, and delegates focused subtasks to fresh subagent instances. It keeps a persistent todo list in~/.xynthis/todo.json that survives sessions, and given a natural-language goal it can build, test, and self-fix a brand-new tool. Safety comes from a per-tool ask/allow policy and an approval flow for risky calls, both on by default. A shadow scan of untrusted tool results for prompt-injection patterns exists as an opt-in extra: it requires a build with --features security-scan (default builds exclude it) plus XYNTHIS_SECURITY_SCAN=1 at runtime. See chatting and the CLI reference.
Computer use
On macOS the agent sees the screen through screenshots plus the accessibility tree, reads pixels with on-device Apple Vision OCR, and acts with real mouse and keyboard events, including clicking elements by their accessibility label. It drives a browser over CDP and targets Safari tabs reliably via AppleScript. Guide-Me mode inverts the relationship: instead of acting, the agent highlights one on-screen target, narrates one instruction, and waits for you to click it. Workflow recording watches you perform a task step by step so the agent can replay it later. See the computer use guide.Voice
A mic button in the app gives push-to-talk input: transcription runs on-device first, with a whisper fallback through the daemon. Spoken requests get spoken replies, a speaker toggle controls read-aloud for typed messages, and Settings lets you pick the voice. See voice in the app.Watchers and automation
Watchers are declarative pollers: point one at a folder or a command’s output and changes flow into the brain as perceptions. Recipes are reusable YAML task templates you run with parameters, andxynthis schedule binds cron expressions to them for always-on automation. A proactive engine checks every 5 minutes for things worth flagging: low disk space, sustained CPU, new Downloads files, long idle. See watchers and automation in the CLI reference.
Skills and learning
Skills are reusable technique docs, optionally with executable scripts, that the agent consults when a matching task comes up. Lessons the agent learns from failures auto-promote into skills, and a learned-action cache replays successful approaches to known tasks without an LLM call. Separately, a rerank head trains on your real recall traffic, so retrieval ranking improves the longer you use it. See skills and learning.Codebase memory
xynthis corpus indexes folders as named collections with AST-aware chunking, BLAKE3 dedup, and auto-watch, so recall can quote your code and docs. xynthis code scan walks a repo and ingests typed structural triples into the knowledge graph, so recall can answer questions like “what calls this function”. See codebase memory.
Models and providers
Fourteen provider presets ship built in, from Anthropic and OpenAI to Ollama and llama.cpp, and the brain’s memory is provider-independent: switch models without losing anything.xynthis auth discover finds every credential already on your machine and wires it into config, and a global --llm flag overrides the model for a single invocation. See model providers and local models.
On-device model
xynthis-llm-serve exposes an OpenAI-compatible /v1/chat/completions endpoint on port 8080 with per-session KV-cache reuse. The local model fine-tunes itself from your brain’s data via LoRA, promoting a checkpoint only when evals pass, and an opt-in readout reports which concepts the model weighed per reply. See the model server, training, and workspace readouts.
The macOS app
The app lives in the menu bar and is the visual face of the brain: chat with a live activity timeline, a session sidebar shared with the CLI, a workspace with terminal, git diff, and web preview panels, and screens for memory, the knowledge graph, and lessons. Status surfaces (a home overview, a Focus Radar screen, a brain dock in the sidebar) keep the brain’s vitals visible, and overlays show exactly where the agent clicks during computer use. An onboarding wizard handles first run, and the app updates itself via SHA-256-verified DMG downloads. See the macOS app.MCP
Thexynthis-mcp bridge gives Claude Code, Cursor, Codex, or any MCP client the brain’s tools: remember, recall, knowledge-graph queries, corpus management, watchers, witnessed-memory confirmation, and a grounded who-is-the-user brief built strictly from verbatim stored facts. Whatever your coding agent stores is there for every other client, because it is all one memory. See MCP integration.