Models are Stateless

Anthropic's memory tool is a concrete example of building memory on top of stateless models — the model still forgets everything between calls, but a tool gives it a place to write and read notes that get loaded back into context.

Anthropic's conversation summarization guide covers what to do once that history grows too long to keep sending verbatim — summarize older turns, keep recent ones intact.

Simon Willison's Catching up on the weird world of LLMs (2025) revisits the basics for a general audience — including the bit about how chat is really just a transcript being re-sent every turn, dressed up to look like memory.