Memory, Reflection, and Long-Horizon Agency
Memory as state
Generative agents
A landmark line of work on generative agents introduced a memory stream, reflection, and planning loop that helps agents maintain consistent, human-like behavior across time. ([nips.cc](https://nips.cc/media/neurips-2023/Slides/72905.pdf?utm_source=openai))
Memory architecture
A practical memory stack usually includes episodic memory, semantic memory, working memory, and task memory, with retrieval conditioned on salience, recency, and relevance.
Memory functions
| Type | Role | Typical failure |
|---|---|---|
| Episodic | Stores experiences | Irrelevant recall |
| Semantic | Stores distilled facts | Knowledge drift |
| Working | Holds current task context | Context overflow |
| Reflective | Synthesizes lessons | Self-confirmation bias |
Open problem
The field still lacks a universally accepted theory of what should be remembered, forgotten, or compressed in long-running agents.
What is the main purpose of reflection in an agent architecture?
Reflection extracts abstractions and strategic guidance from past experience.
Correct answer: To compress raw experience into higher-level lessons and plans
Which memory issue is most directly related to context-window limits?
Working memory is bounded by the model’s context length and prompt budget.
Correct answer: Working-memory overflow
Name one criterion for memory retrieval in an agent.
Other valid criteria include recency, novelty, utility, and salience.
Correct answer: Relevance
Why is long-horizon memory difficult in deployed agents?
This combines utility, privacy, and systems constraints.
Correct answer: Because memories must stay useful, private, and computationally affordable over time