LN: Park et al. (2023) — Generative Agents: Interactive Simulacra of Human Behavior
Bibliographic Reference
Citation: Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. ACM UIST 2023. https://doi.org/10.1145/3586183.3606763 Venue: ACM Symposium on User Interface Software and Technology (UIST) 2023. Stanford University / Google.
Pass 1 — Bird’s Eye View (5 Cs)
| C | Assessment |
|---|---|
| Category | System design + user study |
| Context | Smallville: a 25-agent sandbox where LLM-powered agents live, work, and interact as if human |
| Correctness | Evaluated via expert human raters; ablation studies on memory, reflection, and planning components |
| Contributions | (1) Memory stream: timestamped event log with relevance/recency/importance scoring; (2) Reflection: agents synthesize higher-level insights from memories; (3) Planning: daily schedules decomposed to atomic actions; (4) Emergent social behaviors: election organizing, rumor spreading |
| Clarity | Excellent. Architecture fully described. Video demo available. |
Relevance: ⭐⭐⭐⭐
The memory stream + reflection architecture directly maps to PUMA’s Stage 5 SmartPMO agents, which need persistent memory across PM sessions.
Pass 2 — Content
Architecture: Three Components
Memory Stream ──→ Retrieval ──→ Actor (plans + actions)
↑ |
└──────── Reflection ←─────────┘
↑
(nightly synthesis)
Memory Stream: Every agent event (observation, conversation, action) is stored as a timestamped entry with three scores:
- Recency: exponential decay from the last access
- Importance: LLM-rated 1–10 (“How significant is this?“)
- Relevance: cosine similarity to the current query
Reflection: Every ~200 memory tokens, the agent synthesizes abstract insights: “Klaus Mueller is passionate about environmental issues” from specific memories about Klaus’s actions.
Planning: Agents create coarse daily plans (“Wake up, exercise, work on research, lunch…”) decomposed into 5–15 minute actions.
Emergent Behaviors
- Agents autonomously organized a Valentine’s Day party (never explicitly programmed)
- A rumor about a candidate in the mayoral election spread to 12 of 25 agents within 2 days
- Agents maintained consistent persona traits across multi-day simulations
Ablation Results
| Component removed | Impact |
|---|---|
| No reflection | Agents repeat actions mechanically; no insight development |
| No planning | Agents become reactive; lose time-consistency |
| No memory retrieval | Agents forget context; contradictions emerge |
Pass 3 — Virtual Reconstruction
Q1 (Memory for PUMA SmartPMO): PUMA Stage 5 agents managing PM workflows need memory across sprints. The memory stream architecture (recency × relevance × importance scoring) is directly applicable to a PM agent that must remember: past sprint retrospectives, team velocity patterns, and stakeholder preferences.
Q2 (Reflection for PUMA): A PUMA sprint manager agent could generate nightly reflections: “Sprint 3 consistently underestimates story points for frontend tasks — adjust estimation strategy.” This is Reflexion applied to project management.
Q3 (Scale): 25 agents in Smallville used GPT-3.5/4. PUMA Stage 5 with 5–7 specialized agents (Analyst, Architect, Dev, QA, PM) would require ~35k tokens/day per agent at Smallville scale — expensive with cloud LLMs, feasible with local inference.
PUMA Integration
- Stage 5 SmartPMO architecture: Memory stream + reflection = foundation for PM agent persistence → Smart-PMO-Vision
- Agent design: The importance scoring mechanism for PUMA should weight issues with SLA breaches and P1 priority as high-importance memories
- Multi-agent coordination: The Smallville interaction model (agents talk to each other) maps to PUMA BMAD agent roster conversations → BMAD-Agent-Roster
Related Notes
- PN-MultiAgent-ArchitecturePatterns — social simulation vs. task completion agents
- PN-RAG-Embeddings-VectorDB — embedding-based memory retrieval
- LN-Packer-2023-MemGPT — OS-based LLM memory management
- LN-Shinn-2023-Reflexion — verbal self-reflection mechanism
- BMAD-Agent-Roster