Anthropic engineer: "You don't need better prompts. You need graph engineering: memory that stays, so your agent never forgets anything." In 28 minutes he shows what Anthropic does differently, how to build and structure work with agents. This beats any paid agent course I've seen. Watch it, then read... the graph engineering guide below 👇show more

original link:

Like, I like liking, like likes like likers. But like, this guy like, says like, like a lot.

It is getting ridiculous there. What's next buzzword? You need cosmic rays? There is nothing fundamentally different than using LLMs with scripts and data. :)

The problem is real. I live it every day. The paper’s diagnosis is exactly right. The prescription, for a software project, is where we part ways.

I think I really need this

🙏🏼🙏🏼

This is probably one of the most important shifts in agent design

What does the talk have to do with graphs? I don t see it

I think this is a good share

appreciate you brotha

We need tools to better clean memories. My Claude memory is a hot mess. Would be nice to have tools to prune memory more selectively,

We can now build this ourselves right? Into our own AI agents? I already have the obsidian second brain installed and created, maybe this would boost my agent’s memory even further.

graph engineering is definitely worth diving deep into rn, imo

totally best time investement

Memory that stays is half of it. The other half is that a graph gives you somewhere to put reconciliation. Sixteen agents can each remember perfectly and still hand you sixteen answers. Something has to decide which survives.

The next AI breakthrough may come from better systems around models, not just better models. I read about it in @TheUpsideAI this morning

We do not need bigger models. We need cheaper models. Thanks for your graph engineering, loopsy loopsa, I am pretty happy with my workflow. It works amazingly and is 10x more efficient than without it. Now I need cheaper Sonnet-level models. CHEAPER like less expensive.

never forgets anything is the wrong target. a memory that keeps everything is a memory with no ranking, and retrieval quality drops as it grows. what you want is a system that forgets on purpose and can tell you why it kept the rest.

Coding harnesses shouldn’t have memory

this video was dog shit

@grok 总结下这个视频

This would be an amazing improvement!

the next AI breakthrough might be better memory, not better prompts

exactly

You still need better prompts.

Memory that never forgets anything.

There will be signs. The signs:

It’s mostly a guide on: ✅ Prompt caching ✅ Separating ingestion from retrieval ✅ Low effort for extraction, high effort for reasoning ✅ Batch processing historical data Useful engineering advice, just don’t expect a brand new graph memory architecture inside.

memory beats prompting once agents get real

One thing I really like about this direction is that it shifts the conversation away from "better prompts" and toward better system design.

The evolution of cybersecurity is shaped by breakthroughs in research, market demand, and infrastructure maturity in emerging economies.

Actually the graph engineering is basically the Langraph implementation of nodes

Nice

Better prompts help once. Better memory compounds across every interaction. That’s the real shift from using AI as a chatbot to building agents that improve with context.

pretty insightful

Please sell me your 10x graph engineering course made with leaked $2.2M Anthropic engineer scripts.

Future