During a multi-step task, an LLM agent needs to know what it has already done before choosing its next action. Working memory retains that progress for one task execution; it is transient and limited by the LLM’s context window. section 7.1.2
In a stepwise agent loop, the TaskHandler’s rollout carries prior progress into each step’s system message. After the step completes, its experience is appended to the rollout for the next step. The state is not retained after the task ends. section 7.1.2
Boundary of memory
|
Artifact
|
Derived
from LLM agent experience?
|
Working or
Episodic Memory
|
|
Input
and output of a task execution
|
Yes
|
Episodic
|
|
Lesson
extracted from task execution
|
Yes
|
Episodic
|
|
Rollout
injected into system message
|
Yes
|
Working
|
|
User
preference
|
No
|
N/A
|
|
Skill
|
No
|
N/A
|
|
LLM’s
parametric weights
|
No
|
N/A
|
|
RAG
over Wikipedia
|
No
|
N/A
|
Compare the task rollout classified as working memory with episodic records and artifacts that are not memory. section 7.1.2
In multi-agent systems, a coordinator may discard subagents’ rollouts to avoid filling its context window, while those subagents retain separate long-term memory. section 7.6










