Today’s highlight is a study on how AI agents should remember things, and the rest of the radar ranges from AI-generated video to model cost. I’ve split everything below.

When I first started looking at AI agents from the product side, I thought memory was a volume problem: the more the agent stored, the better.

Anthropic published research on the components of agent memory, and the result points the other way. In one test, with the game Pokémon Red, a model wrote 413 memory files and earned one badge. Another wrote about ten and earned four. The difference was in what each one considered worth storing, such as strategy and plans that the game screen doesn’t show.

Another point that caught my attention: a wrong note in memory repeats in every following session. In the study’s example, a mistaken entry made the agent fail several times at the same spot, until an offline step reviewed the transcripts and fixed the note. Without that review, the agent never got past it.

For anyone building AI products, this changes the conversation. Memory stops being an engineering detail and becomes a product decision: what it is for, what goes into the context, who checks what was stored, and how much it costs to maintain. The study also shows that the same memory prompt behaves differently on each model, so testing before locking in a rule makes a difference. The guides to AI agents, evaluating AI agents and AI agent cost management help structure these decisions.

I see this with enthusiasm. Agents that learn from use and improve over time are exactly the kind of product that creates recurring value, as long as we design memory with the same care we give any other feature.

For those who got curious and want to read the full research, here is the read.

The rest of the radar

  • Gemini Omni 1.2 Flash — 3 to 15 second video with brand references, 30% cheaper than 1.1 (US$ 0.07 per second at 720p); worth re-evaluating the generated-video bill in your product. Google
  • Cheap model with Opus as backup — in a Hightouch test on synthetic data, the hybrid came close to Opus’s accuracy alone for about a quarter of the cost; routing only the hard case is a cost lever. Hightouch
  • Replit and microVMs — Replit swapped containers for microVMs to isolate users’ code; anyone letting an agent run code needs to treat isolation as a product requirement. Replit

That’s it for today. More tomorrow.