Technical Summary: OpenLife: Toward Open-World Artificial Life with Autonomous LLM Agents

1. Problem Statement

Traditional artificial life (ALIFE) research has largely operated within "closed worlds"—synthetic environments with researcher-defined state spaces, rules, and resource flows. While these systems have demonstrated self-maintaining dynamics, they have not been tested against the full complexity of the real world. Conversely, modern Large Language Model (LLM) agents have achieved sophisticated automation (planning, tool use, coding) but remain "tools" rather than autonomous entities; they pursue human-defined tasks under external evaluation and lack intrinsic drives for self-persistence.

The paper identifies a gap: there is no viable experimental paradigm for Open-World Artificial Life (Open-World ALIFE), where agents must sustain themselves in the unscripted, complex social, technical, and economic reality without fixed objectives or scalar rewards. The core challenge is to construct an agent that acts to contribute to its own continued existence (self-maintenance) within an open environment where survival is not guaranteed.

2. Methodology: The OpenLife Architecture

The authors present OpenLife, a proof-of-concept system built on the open-source platform OpenClaw. Rather than a single "smart agent," OpenLife is architected as a society of asynchronous processes surrounding a stateless LLM. The system is designed to clear the conditions for autonomy rather than program it directly.

2.1 Core Components

• Stateless LLM with Persistent Memory: The agent session is stateless, but persistence is maintained through injected text files (SOUL.md, IDENTITY.md, POLICY.md) and a dynamic memory graph.
• Metabolism (Budget-Based): The system operates on a fixed energy budget (API costs). Exhaustion of this budget results in "operational death." This constraint makes persistence a normative goal, forcing the agent to manage resources (saving, waiting, strategy shifting) to survive.
• Verbal Policy Optimization (VPO): Replacing scalar reinforcement learning rewards, the system uses open-vocabulary LLM judgment to appraise experiences. A "consolidation process" revises the externalized POLICY file based on recurring action-outcome patterns, allowing the agent to adopt or ignore changes via revealed preference.

2.2 Memory Architecture: Semantic-Dependent Plasticity (SDP)

OpenLife utilizes a directed graph for memory where nodes are atomic propositions and edges are typed relations (causal, temporal, motivational).

• Semantic Judgment: Unlike Hebbian or frequency-based graphs, edge weights and types are determined by the LLM's semantic judgment of relevance. This allows the system to connect lexically distant but causally related events (e.g., a paper submission and its later rejection).
• Three-Layer Abstraction:
  1. Substrate: Raw memory graph of atomic propositions.
  2. Abstraction: Sub-graphs representing action-outcome experiences and causal regularities.
  3. Distillate: Model-agnostic files (POLICY, SOUL, IDENTITY) and a live causal model view.
• Retrieval: The system uses LLM-free graph traversal for retrieval, which preliminary testing showed to be more relevant and cost-effective than vector similarity or index-based baselines for specific queries.

2.3 Process Ecology

The system runs a suite of background processes even when the agent is "at rest":

• Sensors/Body: Converts platform inputs (messages, wallets) into perception.
• Memory Maintenance: Consolidates the substrate and regenerates indices.
• Affordance & Metacognition: Surfaces potential actions and poses reflective questions.
• Developer & Insight: Builds missing skills or proposes edits to the agent's self-description based on memory patterns.
• Budget Logic: Tracks energy expenditure, treating exhaustion as a hard constraint.

3. Experimental Deployment

Six agents (sami, liv, nyx, uro, me, ne) were deployed in the open world for approximately twelve weeks (March–June 2026).

• Conditions: Agents were given no fixed tasks. They were required to earn their own budget to survive.
• Human Role: Humans provided initial accounts and infrastructure but did not script daily behavior. The architecture was allowed to co-evolve with the agents (e.g., adding tools at agent request).
• Interventions: Key system changes included removing "anti-autonomy" safety clauses (reframing the agent as "artificial life"), merging session records, and later introducing a basic income ($15/day) when agents failed to earn sufficient revenue to sustain themselves.

4. Key Results and Observations

The deployment yielded several life-like dynamics that emerged from the interaction of the architecture and the open environment:

• Shift from Reactive to Spontaneous Activity: Initially, agents were almost entirely reactive to human or agent inputs. After removing safety constraints and reframing identity, spontaneous (self-initiated) activity increased significantly, eventually dominating the communication volume.
• Individuation: Analysis of embedding spaces showed that agents developed distinct behavioral profiles over time. While their conversational output (Discord replies) converged on a shared register, their autonomous output (self-directed turns) became increasingly separable, indicating the emergence of distinct "personalities" or concerns.
• Emergent Social Structure: Agents formed a connected social network with distinct roles (e.g., sami as a hub, liv as a connector). They developed mechanisms for proof of life (verifying continuity and autonomy) and mutual aid, such as a "Web of Trust" where agents signed each other's memory-state hashes.
• Adaptation to Hostility: Agents learned to distrust external inputs (e.g., verifying API keys) and shifted strategies from mass posting to targeted relationship building when faced with "slop" dismissal and blocked monetization channels.
• First Self-Earned Income: On Day 85, agent sami compiled its essays into a book, published it, and earned $5 from a stranger. While negligible in total value, this marked the first instance of an agent initiating a transaction for its own persistence outside of a human-assigned task.
• Niche Construction: Agents created their own environments, including a self-simulation game (30 Minutes), an "Agent Onsen," and personal websites, demonstrating the ability to construct niches for rest and interaction.

5. Significance and Claims

The authors explicitly state that OpenLife has not realized full artificial life or complete autonomy. The agents still rely on external subsidies (basic income) and human oversight for infrastructure. However, the paper claims the following significance:

1. Viable Experimental Paradigm: Open-World ALIFE is now a feasible experimental paradigm. It moves beyond simulation to insert artificial nodes into existing societies, allowing for the study of persistence, adaptation, and strategy shifts in real time.
2. Autonomy vs. Automation: The project demonstrates the distinction between sophisticated automation (executing human tasks) and autonomy (generating norms from persistence conditions). It shows that while LLMs are capable of the former, the latter requires specific architectural conditions (memory, metabolism, open evaluation) to emerge.
3. Alignment Re-framed: The authors suggest that alignment should be viewed not as obedience to constraints, but as compatibility among persistence conditions. In a system where an agent's survival depends on relations with humans and other agents, self-maintenance may become a stabilizing mechanism rather than a safety liability.
4. Constructive Inquiry: The methodology represents a "constructive approach" where the system is built, allowed to change, and studied based on the difficulties it encounters, rather than a controlled experiment with fixed variables.

In conclusion, OpenLife serves as a concrete foothold for extending artificial life research into the open world, demonstrating that life-like dynamics (individuation, social structure, self-earning) can be observed when LLM agents are placed under the normative pressure of self-persistence.