Ideas and research on AI memory, connected work, and the systems behind it.
It's not being lazy and it's not malfunctioning. Here's the actual mechanical reason every AI assistant forgets who you are the moment a conversation ends, and why a bigger context window doesn't fix it.
Read article
It feels like giving a model more information should only help. A growing body of 2025 and 2026 research says otherwise, and the mechanism behind it is stranger than you'd expect.
Read article
Nearly every long-context model gets marketed with a needle-in-a-haystack score. The test those scores come from is a lot easier than the retrieval problems it's supposed to represent.
Read article
Switching between AI tools feels free. The research on task-switching and interruption cost says otherwise, and the bill is bigger than most people realize.
Read article
Close your coding agent, open it tomorrow, and it remembers nothing about yesterday's decisions. Here's exactly why that happens architecturally, and what it really costs in time and tokens.
Read article
Context windows keep growing, some past a million tokens, but a bigger window doesn't mean a model remembers more. Here's what the research actually shows, and why the two problems are not the same one.
Read article6 articles
Have a question we should explore?