Keeping a long conversation short

A session only grows. Every turn is sent to the model on every later turn, so a long conversation gets slow and expensive before it gets confusing.

Example

print("after three turns:", len(await session.get_items()), "items")

removed = await session.pop_item()
print("removed the last one:", removed["role"] if isinstance(removed, dict) else removed)
print("now:", len(await session.get_items()), "items")

await session.clear_session()
print("after clearing:", len(await session.get_items()), "items")

Taking the last one back

python

removed = await session.pop_item()

pop_item removes the most recent item and hands it to you. The obvious use is an undo button: the customer sends something by mistake, you take it off the end and the model never sees it again.

Starting over

python

await session.clear_session()

Everything, gone. Useful when a conversation is finished, or when a customer asks you to forget what they said, which is a request you will eventually have to honour.

The three ways to keep it short

Approach	Costs you	Use when
Let it grow	more tokens every turn	short conversations, and only those
Pop the old items	the early turns are gone	the beginning stopped mattering
Summarise, then clear	one model call to write the summary	long conversations where the early turns still matter

The SDK gives you the first two. The third is yours to write: read the items, ask an agent to sum them up, clear the session, and put the summary back as the first item.

It fails at the worst time

This is not a problem you notice in testing. It arrives as a context length error in production, on your longest running conversation, which is usually your most important customer.

Try it yourself

Pop twice and print what comes back each time.
Write a loop that keeps only the last four items.
Clear the session and run one more turn to confirm it starts fresh.

PreviousRemembering the conversation

Every expert started right here.