Episodic memory lets an LLM agent retain discrete experiences from completed task executions and recall relevant ones for future tasks. It can help the agent avoid repeated work and apply lessons from earlier runs. section 7.0 section 7.1.2
An episode is one unit of experience. Memory records can draw on task inputs, results, tool calls, trajectories, or lessons extracted from an execution. What to retain is a design decision that can affect performance; it is separate from how records are stored. 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 experience-derived episodic and working memory with artifacts outside this definition of memory. section 7.1.2
Working memory tracks progress within a single task, typically through a rollout supplied to the model. It is transient and limited by the context window. Episodic memory instead carries experience across tasks; recalled information must be supplied to the stateless model at inference time. section 7.1.2 section 7.7
A task lifecycle can incorporate episodic memory in three steps:
- At the start, recall relevant Episode objects for the current task. section 7.7
- Provide the recalled experience in the model’s prompt or context so it can influence execution. section 7.7
- At the end, record the completed execution as an Episode for later recall. section 7.0 section 7.7
Memory is the core type responsible for recording and recalling Episode objects through a BaseMemoryStore backend. Recall and storage choices are coupled: semantic-similarity recall, for example, requires vector-based storage. section 7.7
Memory across subagent dispatches
A subagent configured with episodic memory can preserve context across separate dispatches, even when each dispatch starts with a fresh execution rollout. This does not resume an interrupted task: without memory, the main agent must supply necessary prior context in the new task instructions. section 9.2.3

