简体中文 · English

Generative Agents: Memory, reflection and planning¶

Implements relevance/recency/importance retrieval from a memory stream, reflection triggered by cumulative importance, and planning from retrieved memories.

Paper information¶

Field	Details
Paper	Generative Agents: Interactive Simulacra of Human Behavior
Institution	Stanford University / Google Research
First public release	2023-04-07
Author code	Public author code.
Local adapter or method key	`generative-agents`
Local reproduction code	`src/auto_research/agent_research/`

Original paper summary¶

Background and main changes¶

Putting all history in an LLM context does not sustain coherent long-term behavior. Store observations in a memory stream and retrieve by recency, importance and relevance. Once cumulative importance crosses a threshold, generate higher-level reflections, then use memory and current state for schedules and action plans.

flowchart LR
    O["Observation"] --> M["Memory stream"]
    M --> R["Recency, importance and relevance retrieval"]
    M --> T["Importance threshold"]
    T --> F["Higher-level reflective memory"]
    F --> M
    R --> P["Planning"]
    P --> A["Action or new observation"]
    A --> O

Key figure from the paper¶

Figure 5: Core architecture, modules and connections. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\[ s(m,q)=\alpha\,\mathrm{recency}(m) +\beta\,\mathrm{importance}(m) +\gamma\,\mathrm{relevance}(m,q). \]

Offline and online results¶

The 25-agent Smallville experiment demonstrates invitation diffusion, relationship formation and coordination. Human evaluation and ablations find that observation, planning and reflection all affect believability. No production A/B.

Local reproduction¶

EvoMem-mini, 120 episodes, seed 42, memory size 24. Retrieve top 3 by the three-part score and synthesize reflections from cumulative importance.

Metric	Long-context baseline	Generative Agents
Joint success	1.0000	1.0000
Average cost	64.5000	1.7900
Retrieved memories / reflections	0 / 0	354 / 30

auto-research agent-eval --method generative-agents --benchmark evomem-mini \
  --episodes 120 --memory-size 24 --seed 42

Stable metrics: p0-missing-agent-mini-suites-seed42.json.

Scope¶

Reproduces memory scoring, reflection triggers and planning dependencies. No 25 LLM personas, Smallville map or human believability evaluation were run. Compare only execution success, proxy context cost and internal state.