State, Memory and Context Over Many Steps

You will be able to

• Place information in internal knowledge, short-term memory (the context) or long-term memory by how often it's needed.
• Explain why a growing context fails twice: by overflowing the window, and by being used less reliably as it grows.
• Show why dropping the oldest messages can drop the task itself.
• Pin what must stay in the context and summarise (compact) the rest, testing what the summary keeps.
• Keep exact or structured information as harness state outside the context, and load details just in time.
• Write handoffs to sub-agents or later sessions that carry the goal, inputs, limits and gaps.
• Add, merge, replace and delete memories, and choose how to resolve contradictions.

Topics

1. 1Three kinds of memory
2. 2The context fills up
3. 3Dropping old messages
4. 4Pin what matters, summarise the rest
5. 5State outside the context
6. 6Handoffs
7. 7Managing memory, and wrap-up