Most AI tools are reactive: you ask, they answer. But what if an AI could sit in the background, watch your work, and surface insights before you even think to ask?
That was the premise behind Mnemon — an always-on agentic daemon I built as part of my own toolchain. It runs as a background process, maintains a persistent memory graph, and uses idle CPU time to do "background thinking" — synthesising observations, updating goal states, and queuing actions for later review.
The core architecture is deceptively simple: an event loop that cycles through three phases — observe, think, act. Each phase has a configurable budget (wall time + token spend). The daemon subscribes to OS-level events via IPC: file system changes, clipboard updates, terminal output, even inactivity periods.
One of the trickiest parts was managing autonomy safely. The daemon should be helpful but not rogue. I implemented a simple autonomy slider (0–1) where lower values require human confirmation for any action, and higher values allow the daemon to act unilaterally within a defined sandbox.
The memory layer uses a hybrid approach: a vector store for semantic recall, a graph store for relational facts, and a rolling timeline for recent events. This lets the daemon answer questions like "what was I working on three days ago?" with surprising accuracy.
If you're thinking about building your own AI daemon, start with the smallest useful loop: observe one thing, think one thought, surface one suggestion. Complexity can always come later.