Skip to content

Summary: Park et al. (2023) — Generative Agents

Full citation: Park, J.S., O'Brien, J.C., Cai, C.J., Morris, M.R., Liang, P., and Bernstein, M.S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. 36th Annual ACM Symposium on User Interface Software and Technology (UIST 2023).

Key Contribution

This Stanford/Google paper demonstrates a 25-agent social simulation where LLM-powered agents inhabit a sandbox environment (Smallville), maintain persistent memories, form relationships, coordinate activities, and produce emergent social behaviours. The paper's primary contribution to this wiki is the memory stream architecture — the most detailed and influential design for agent episodic memory, directly relevant to memory-architectures.

Memory Stream Architecture

Each agent maintains a memory stream — a chronologically ordered list of natural language observations recording everything the agent perceives. Each memory entry includes a timestamp, a natural language description, and metadata.

Three-factor retrieval scoring determines which memories are injected into context at each reasoning step:

  1. Recency: exponential decay function — recent memories score higher. This ensures the agent's current reasoning is grounded in its latest observations.
  2. Importance: each memory is assigned an importance score (1-10, generated by the LLM) reflecting its significance. Mundane observations (eating breakfast) score low; life-changing events (discovering a new relationship) score high.
  3. Relevance: cosine similarity between the current query/situation embedding and each memory's embedding. This surfaces memories that are semantically related to the current context.

The final retrieval score is a weighted combination of all three factors. This multi-factor approach is more sophisticated than pure semantic search (which misses temporal context) or pure recency (which misses important old memories).

Reflection: Memory Abstraction

Periodically, agents generate reflections — higher-level abstractions synthesised from recent memories. When the sum of importance scores since the last reflection exceeds a threshold, the agent asks itself questions like "what have I learned about [topic]?" and generates summary statements that are themselves added to the memory stream.

Reflections serve as a form of memory compression that preserves meaning: instead of losing detail through truncation or summarisation, the agent generates new higher-level entries that coexist with the detailed observations. This connects to context-management-risks — reflection is a structured alternative to the lossy compression strategies described there.

Planning and Action

Agents generate daily plans (high-level schedules) that are recursively decomposed into hourly and sub-hourly actions. Plans are stored in the memory stream and can be revised when new observations conflict with the planned activity. The planning mechanism creates goal-directed behaviour over extended periods.

Limitations for Safety-Critical Applications

All 25 agents shared the same base model (GPT-3.5-turbo). The emergent social behaviours (information spreading, relationship formation, coordinated party planning) are impressive but do not demonstrate epistemic-independence. Every agent's reasoning is generated by the same model with the same distributional biases. Emergent macro-level behaviour is not the same as independent assessment — a key distinction developed in the multi-agent-taxonomy page.

Sequential discrete time steps, not concurrent operation. The simulation advances in discrete steps; agents do not operate simultaneously. This is Pattern 0 (single-agent role simulation) in the taxonomy, not Pattern 9 (concurrent multi-agent).

Relevance to This Wiki

The memory stream architecture is the most influential design for LLM agent memory and directly informs memory-architectures. The three-factor retrieval scoring (recency + importance + relevance) provides a concrete implementation for the episodic memory concept. The reflection mechanism offers a structured approach to the context-management-risks challenge — generating higher-level abstractions rather than lossy compression. For safety-critical applications, the architecture demonstrates both the potential (persistent, structured memory across extended sessions) and the limitation (all agents sharing a single model cannot provide independent verification).