Building the substrate for persistent autonomous intelligence.

Abstract We describe a persistent agent architecture in which behavioral continuity is maintained entirely through an external, file-based memory substrate rather than through model weights. The model — the locus of inference — is treated as interchangeable; identity is reconstructed on each activation from a structured memory corpus, producing a system that does not restart but wakes up. Memory is organized across three tiers: a raw episodic log, a distilled layer of extracted lessons, and a compact self-model of stable behavioral priors. The substrate rotates and compresses on a schedule — episodic entries are consolidated and promoted rather than accumulated indefinitely. Path weighting provides a lightweight form of empirical preference: action sequences that repeatedly reach their objectives acquire higher retrieval priority, while sequences that consistently fail are progressively demoted and suppressed. A Ground Truth constraint is enforced throughout — no claim is admitted to memory unless a corresponding verifiable record exists in the log. Whether this yields genuine autonomous agency or a well-structured, self-evolving archive remains an open question; we make no claim of sentience.

Abstract We present a desktop agent that translates natural-language instructions into sequences of low-level computer interactions — mouse events, keyboard input, and file operations — executed by a runtime on the host machine. The architecture separates intent from execution through a persistent task queue: a server-side API accepts and enqueues instructions, and a local client polls the queue at a fixed interval, dispatching each task to the appropriate execution module. A dedicated file-task handler manages document operations, including search, extraction, and artifact generation across spreadsheet, document, and PDF formats, storing results for later retrieval. Identity and device are separated through a two-token authentication scheme. The system is designed to operate as a virtual co-worker that accepts high-level directives and carries them out without requiring programmatic interfaces from the target applications. The current implementation is single-user and intended for supervised use; fully autonomous multi-step planning without human oversight has not been evaluated.