Everyone in AI is talking about agent memory the way people at university talk about sex: constant chatter, very little practice. Here’s what I’ve been building, teaching, and arguing about while the noise gets louder.

Frontier labs quietly settled the argument the industry is still having

The industry spent two years arguing about context window size. The frontier labs have already moved on to context window management: what to keep, what to distil, what to retrieve, and when.

Read OpenAI’s Astra release closely and the memory patterns are hiding in plain sight.

When Astra “keeps notes across context windows,” that’s the scratchpad pattern, context offloading, detail leaving the token-limited window for a store the agent pulls from on demand. When “earlier context windows remain searchable,” that’s working memory the agent can query mid-task.

And when the model is “trained to specifically pull only the context that matters,” retrieval judgement is moving into the weights, not bolted on as a tool.

The line most people skipped is the one that matters: each compaction “can leave out details about why a fix failed.” That’s the failure mode every engineer knows, an agent burning a fix it already tried. If those implications weren’t obvious to you reading the release, that gap is exactly what I train people to close.

Memory, not intelligence, is the mode now

Intelligence from the model is basically a commodity at this point.

Memory is the moat, and everyone knows it.

AWS, Microsoft, Oracle, Databricks, the frontier labs, the startups, all racing to own it.

The trouble is that a lot of platforms can talk about agent memory rightly, but few can talk about it credibly.

Those are different things.

Credibility comes from solving the problem with something that outlives the article and the slide deck.

At my day job at Oracle that means two things for me: educating developers wherever we can, and shipping something they can actually run, pip install oracleagentmemory.

What I keep teaching is that memory is not one thing.

Humans have working memory, episodic, semantic, procedural, held in different parts of the brain, and agents need the same division of labour.

I organise the whole discipline around four Rs:

• Retain
• Recall
• Reuse
• Refine

Plus one principle people resist: you don’t delete, you forget.

Build for forgetting and you’ll see every piece of information differently.

The people who survive this stay in the arena

I’ve also come to believe enablement designed from a back office is theatre. Whoever delivers it has to stay on the front line, see what lands with real customers, then scale it.

I’d rather train twenty specialised AI architects with a growth mindset than push a thousand people through a vague “ramp up on AI” objective.

The twenty will teach the thousand anyway.

The same logic applies to how you show up with customers.

Being conversation-ready is table stakes now; everyone can hold a chat about token maxing or continual learning.

The real climb is to proof-ready, showing working code and reference architectures, then forward-looking, telling a customer where the field goes six months out and being right often enough that they call you by name.

That last level isn’t a sales tactic. It’s earned by building in public and being genuinely early, the way I bet on agent memory three and a half years ago.

The one thing I’m deliberately doing slower

I’ll admit something. I’m offloading more cognition to AI than I’d like, sometimes just asking it to give me the steps so I don’t have to think. If everyone points the same models at the same problems, we all arrive at the same solution shape, and the differentiator quietly disappears.

So this era, I’m choosing a few tasks to do the slow way, reading the paper myself, whiteboarding, keeping the muscle alive. Spend some time this week learning something without AI.

That’s coming from someone who ships an AI learning tool.

If you want to stop nodding along in memory conversations, the waitlist for my Maven course, Introduction to AI Agent Memory, is open: https://maven.com/ascxend/introduction-to-ai-agent-memory​