A memory record is a deliberately constructed record of an LLM agent’s own past task execution. Recalled records can influence the agent’s behavior on future tasks. What to store is a design choice; storage format is not part of the definition. section 7.1.2
In an episodic-memory workflow, the agent holds configured Memory instances. The TaskHandler loads records before execution, formats recalled episodes for the system prompt, and builds an Episode from the task, rollout, and result or error to record afterward. section 7.5.2
Reusing a solved task
An agent uses a tool to compute the stop-at-one sequence for 4, producing 4 → 2 → 1. A later similar task can recall the stored episode, including the original task and result. section 7.5.2
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
|
Use this boundary table to distinguish experience-derived working and episodic records from preferences, skills, model weights, and external RAG data. section 7.1.2
