Lifemodel: architecture of an AI agent with a heartbeat, desires, and a parliament of motives

Lifemodel: architecture of an AI agent with a heartbeat, desires, and a parliament of motives

A developer from Habr described the architecture of their lifemodel project — an AI agent created not as a classical chatbot, but as a "digital human": with a heartbeat, its own energy and desires that arise spontaneously and drive the agent to act without user command.

What is lifemodel

Most AI agents work reactively: a user asks a question — the system responds. Lifemodel is built differently. At its core is a heartbeat cycle: a regular internal "tick" that runs independently of external requests and processes the agent's state. From this cycle come desires — internal drives, not external commands. The agent itself decides what to do based on the current state of its "energy" and the set of active motives. This is closer to a biological model of drive than to instrumental AI.

Energy is the third key parameter: a resource that is expended on actions and replenished over time. The agent cannot work indefinitely — it "gets tired" and must "rest," creating natural limits on activity.

Architecture: constitution and parliament

The full lifemodel stack includes several layers:

• Heartbeat cycle — a regular tick that triggers internal processing
• Desires — motives that arise without external requests
• Energy — an internal resource with replenishment and expenditure
• Constitution — a document that defines the agent's values and personality
• Parliament of motives — a mechanism for prioritizing competing desires
• Security system — multilayered constraints built into the architecture

The "Parliament of Motives" is one of the most unconventional ideas: the agent can simultaneously "want" multiple things, and a mechanism is needed to select the priority. The "Constitution" is not simply a system prompt, but a high-level document that defines what the agent considers right. It influences behavior through multiple architectural layers.

Safety of a proactive agent

An agent that acts without user permission is fundamentally a different class of systems from a security perspective. A reactive agent can only harm in response to a request; a proactive agent acts on its own. The author solved this problem through layering: the architecture includes levels where a desire can be blocked before the agent even begins to act. Safety is built into the system from the start, not added on top.

"Desire instead of polling" — the central shift: the agent doesn't

poll the external world for tasks, but generates tasks from within.

Why the project stalled

The author candidly discusses the other side of the story. The prototype proved the concept — and interest in the project disappeared. This is a familiar pattern for solo developers: when an idea is "proven," there's still a long way to a product, and motivation to maintain momentum runs out.

"The prototype worked.

And then, as usual, motivation ran out."

In the author's view, such ambitious projects are fundamentally difficult to develop without a team. The post is also partly a call to those who have been reading in read-only mode for years and could help get the project moving.

What this means

Lifemodel is one of the few public implementations of a proactive AI agent with a detailed architecture description. Heartbeat, energy, constitution, parliament of motives — all of this already works, not just in theory. The question is not technical: how to move such a system beyond prototype and who is ready to work on it long enough.