Working memory, also called short-term or contextual memory, records the current user-agent session: earlier turns and actions taken by the agent. It gives separate LLM calls conversational continuity, even though the model itself is stateless. section 1.5.4 section 5.1.1
The session lives outside the model in the agent harness or application. For each request, the harness sends a selected view of the history; anything omitted is invisible to the model. Context-window and token-budget limits make curation necessary as a conversation grows. A policy can trim older turns, summarize them, or combine both approaches. section 5.1.1 section 5.5.3
A travel preference
A user requests evening flights. The agent can continue applying that preference several turns later if the relevant session history is included in the context. To use the preference in a new session weeks later, the application must first save it to persistent long-term memory and retrieve it when relevant. section 1.5.4
Working memory preserves what was said and done; state tracks structured operational variables such as goals, progress, and collected parameters. Long-term memory persists useful information across sessions. Frameworks may combine or separate these abstractions, so their object names do not always match the conceptual boundaries. section 5.1.3 section 1.5.4
Table 5.1 Names given to short-term memory, long-term memory, and state in the main agent frameworks
| Framework | Short-term memory | Long-term memory | State |
LangGraph |
Short-term memory |
Store |
Graph state |
LlamaIndex |
Chat memory buffer |
Memory blocks |
Workflow context |
Letta |
Message buffer |
Archival memory |
Core memory blocks |
OpenAI Agents SDK |
Session |
Not built in |
Run context object |
Google ADK |
Session events |
Memory service |
Session state |
Compare the names used for session history, persistent memory, and state across agent frameworks. section 5.1.3
Listing 5.4 Illustrative Python excerpt showing a custom session that curates short-term memory before each agent run
from agents import Agent, Runner
from agents.memory.session import SessionABC
from collections import deque
from typing import List
class SummarizingSession(SessionABC): #1
def _materialize(self):
trimmed = self._trim_to_last_turns(list(self._items)) #2
if self._summary:
summary_msg = { #3
"type": "message",
"role": "assistant",
"content": f"Running summary:\n{self._summary}",
}
return [summary_msg] + trimmed
return trimmed
async def _refresh_summary_locked(self): #4
prompt = f"Update the summary with older history:\n{drop_text}"
result = await Runner.run(summarizer, prompt)
self._summary = (result.final_output or "").strip()
agent = Agent(name="Support", model="gpt-5",
instructions="Troubleshoot step-by-step.")
session = SummarizingSession("support_demo", max_turns=6) #5
result = await Runner.run(agent, user_input, session=session) #6
Inspect a schematic OpenAI Agents SDK session that keeps recent turns and prepends a running summary; the excerpt omits plumbing and is not runnable as printed. section 5.5.3










