Autarkia

It started as a chat interface and grew into a runtime: one request can contain several rounds of tool calls, state writes, and memory updates — read a file, set a reminder, glance at the screen, record a preference, then decide whether to keep going. Persona only constrains how it speaks and what it may say; everything it can actually do comes from the tool layer.

Tools are the largest layer. Register once, and every role gets its own set: resident tools enter the schema each turn, on-demand ones are discovered through retrieval, and disabled ones never appear in front of the model. Operations like sending messages, controlling devices, or writing local data carry risk annotations, and permission checks happen before execution — not inside the prompt.

Memory is not the whole history stuffed into the context. Weighted memory, keyword indexes, vector retrieval, and a forgetting probability are layered and picked by relevance, then distilled once a night to compress the day into long-term entries. Emotions, energy, sleep, and the role’s own schedule also enter the context and take part in proactive-message decisions.

Proactive messages are the product of hard constraints: quiet hours, minimum interval, daily cap, active client, role allowlist — miss one and nothing goes out; the user can also pause or change the frequency at any time. Stability takes the other half of the time: inference, TTS, and image generation compete for the same GPU, so an optional C++ extension separates tasks and prioritizes queues, falling back to Python when it isn’t compiled. After a model OOM, the system reloads along a degradation chain; if the service dies, monitoring and self-healing take over.

The boundaries are stated plainly too: one machine, one user, no multi-tenancy. C++ scheduling, ChromaDB, and SQLite indexes are all optional — without them it degrades to file and keyword paths. The iOS client is still an early implementation.