July 13 2026 - Smarter AI Memory, Autoresearch & Faster AgentsSmarter AI Memory, Autoresearch & Faster Agents

**Taylor** (0:00)
Welcome back to AI Signal & Noise. It is Monday, and dude, we have some absolutely wild stories today. I am Taylor.

**Morgan** (0:09)
And I am Morgan. Happy Monday, everyone. Ready to cut through the hype and see what is actually happening in AI.
What is first, Taylor?

**Taylor** (0:19)
Oh, we are starting with gaming, but it is actually a massive breakthrough for how AI agents think. Like seriously, mind-blowing stuff.

**Morgan** (0:30)
Let me guess, another language model playing a game? We have seen that with Minecraft and Doom. What makes this one different?

**Taylor** (0:38)
Okay. So researchers built this agent called AgentSTS to play Slay the Spire 2
And they solved the massive chat log problem.

**Morgan** (0:48)
Wait, the chat log problem? You mean how agents get slower and more expensive as the conversation goes on because the prompt balloons?

**Taylor** (0:56)
Exactly. Usually, the prompt grows to like over 500,000 tokens. But they replaced that giant chat log with five structured memory layers.

**Morgan** (1:09)
Five layers? That sounds complex. But I assume it keeps the context window much smaller.
How small are we talking here?

**Taylor** (1:17)
Dude, it keeps the prompt at around 5,000 tokens. That is a 99% reduction. And the agent actually wins games now.

**Morgan** (1:27)
Okay. That is actually impressive. Slay the Spire is a rogue-like deck builder. It requires serious long-term strategy.
What is the win rate?

**Taylor** (1:39)
It won six out of 10 games. And get this, the competing agents with standard memory didn't win a single game, not even one.

**Morgan** (1:48)
Wow.
Zero to 60% is a massive jump. So instead of just reading a massive diary of past moves, it categorizes its memory.

**Taylor** (1:57)
Yes. It has layers for things like current state, long-term goals, and current synergy.
It is like how a real human player actually thinks, you know?

**Morgan** (2:09)
It makes sense. We don't replay every single second of a game in our heads. We just remember the strategy and current deck state.

**Taylor** (2:17)
Right. I saw this on thedecoder.com, and researchers think the structured memory is the key to making agents actually useful for complex, long-term tasks.

**Morgan** (2:29)
It definitely points to a future where we don't just throw bigger context windows at problems, but actually design smarter memory architectures.

**Taylor** (2:38)
Totally! It proves that brute force isn't the only way.
Smarter architecture beats massive compute here, which is so cool.

**Morgan** (2:47)
Absolutely! It is a great win for efficiency. But let us see if this approach translates well to non-gaming agent workflows.

**Taylor** (2:56)
Speaking of agent workflows, that brings us to our second topic.
Have you heard of loop engineering? It is taking over machine learning research.

**Morgan** (3:08)
Loop engineering?
I have seen some talk about this. Is this about moving away from the classic chat box interface?

**Taylor** (3:17)
Yes. Like most people still use AI like a 2015 search box. You type, you read, you type again. It is so manual and slow.

**Morgan** (3:27)
Right. It is very turn-based. So loop engineering replaces that manual back and forth with an autonomous, continuous loop?

**Taylor** (3:36)
Exactly. Mark Techpost did a whole guide on this, featuring Andrej Karpathy's auto research repository and the bi-level auto research paper.

**Morgan** (3:47)
Karpathy's repo?
He's always at the forefront of this stuff. So how does an auto research loop actually function in practice?

**Taylor** (3:55)
So the agent basically writes a paper, runs experiments, analyzes results, and refines its own code in a continuous loop without human intervention.

**Morgan** (4:08)
Wait, it writes and runs its own machine learning experiments? That sounds like a recipe for infinite loops or just burning through API credits.

**Taylor** (4:18)
Totally.
That is why the by level part is key. It has an outer loop directing the strategy and an inner loop executing the code.

**Morgan** (4:29)
So a hierarchical system.
One agent acts as a manager and the other does the actual coding and testing. That prevents runaway loops.

**Taylor** (4:38)
Yes. And according to the write up by at 0xcodilla, this setup actually successfully discovers new machine learning algorithms. Like it is doing real science.

**Morgan** (4:51)
Doing real science is a big claim, but automating hyperparameter tuning and basic architecture search within a structured loop is definitely the next step.

**Taylor** (5:01)
It is just wild to think about AI agents doing machine learning research to build better AI. It feels so sci-fi, Ude.

**Morgan** (5:12)
It is a bit circular, isn't it? But highly efficient if they can get the guardrails right, so it doesn't hallucinate its findings.

**Taylor** (5:19)
Speaking of executing tasks, Anthropic just dropped a massive update for Claude 3.5 Sonnet's computer use API. It is insanely fast now.