#16: AI With Me by Richmond Alake - OpenAI’s best memory feature survived five days
|
One of the exciting model release of the year happened last week, when OpenAI released GPT-6 Astra. And also at the same time we saw claims of AGI re-emerge. AGI or not, today, I want to walk you through the most instructive product failure I’ve seen this year, because it quietly proved a thesis I’ve been arguing for over a year. The most important paragraph in the Astra release never said the word memory In fact the whole article never even mentioned the word 'Memory' once When OpenAI shipped GPT-6 Astra, most people focused on the coding benchmarks. I focused on a single paragraph about Codex.
"With Astra, we’re introducing a new way for Codex to preserve and retrieve context when the context window fills. Historically, models have used compaction to summarize work during long sessions, such as when debugging complex issues or tackling large refactors. Each compaction can leave out details about why a fix failed or how a component behaves. In Codex, Astra can keep notes across context windows, preserving accumulated details without repeatedly compressing them into a single summary. Earlier context windows remain searchable, so Astra can find requirements or test results from previous messages and tool outputs—even if that information wasn’t captured in its notes."
Read closely, and the design reveals itself. “Why a fix failed” is workflow memory, a type of procedural memory. “How a component behaves” is entity memory. Keeping notes across windows is working memory, and making earlier windows searchable is the retrieval side of it. These are the exact patterns I teach, and none of them are research problems. They’re engineering. In fact, you can join me this Friday, as we will explore some of this patterns. The story doesn't end thereDays later, OpenAI switched it off. Thibault Sottiaux, the Codex product lead, posted that the opt-in experiment caused early stops and replies to older messages, affecting roughly four to five thousand users. Disabled, with a usage reset by midnight. Neither symptom was a model failure Look at what actually broke. Replies to older messages means the recall tier beat the live message. Earlier windows stayed searchable, and the model started treating a retrieved turn as the current instruction, with nothing weighting “what you just said” over “what I can find.” Early stops came from a handoff requesting a fresh window before the notes finished saving. Again, we will discuss this more in the webinar. This Friday at 5pm UK, I’m building the Astra pattern from scratch, live, with the fixes baked in, and sharing the full code with everyone. Register at the link in the comments. |