Brilliant piece Jeff. The memory moat thesis is spot on, but there's an interesting twist emerging:
The data we are observing is clear - users who experience memory-enabled AI show dramatically higher retention and engagement. But paradoxically, the strongest moat might not come from locking memories inside one AI assistant.
Think of memories like your digital consciousness - you'll want them to flow seamlessly across different AI tools and apps, just like your thoughts aren't confined to one conversation. The Model Control Protocol (MCP) trend signals this future.
The real moat will be built by whoever makes memories both powerful *and* portable. Users who experience the magic of AI that truly knows them won't go back - but they'll demand control over their digital minds.
Memory isn't just a feature, it's fundamental infrastructure for the AI era. And users will gravitate to systems that enhance their agency rather than diminish it.
Hmm. I went from asking myself why OAI would bother making their memory portable. But on 2nd thoughts, (i) they can charge rent for the privilege and (ii) they can use to to uniquely gain access to even more memory using a MCP like tag along.
Totally agree. Memory is the real unlock. It’s core to the product thesis I’m thinking through. Curious how long until it stops feeling like a feature and starts feeling like the baseline.
I totally agree with you! Memory is becoming the deepest moat in AI, and that’s exactly why we’re building Glasp (glasp.co) — to capture and preserve the knowledge that shapes us. As memory shifts from a passive record to an active, personalized layer, tools like Glasp will play a critical role in helping people reflect, connect, and learn over a lifetime. 🚀
“Work goals. Relationship doubts. Health concerns. Every passing thought, every quiet worry.”
All mined and monetized, sold to whomever can pay. Those work goals? Your employer has them. Health concerns? Your health insurance company is on line 2 about your rate increase. Every private though, every quiet worry, used against you to tailor everything from targeted ads to phishing attempts.
And before you call me a Luddite or paranoid, realize I work on GenAI, I’m the call from inside the house, I KNOW what they’re planning because I’m part of those meetings.
Memory is definitely the unlock for personalization—agreed. But I'd push back on where memory actually belongs in the stack.
The current approaches—RAG retrieval, prompt-injected profiles, "episodic and semantic memory" represented as text blurbs—feel like we're cargo-culting neuroscience terminology without understanding the underlying mechanisms.
A few problems:
Memory isn't a model dimension. You can't embed personalization into weights and expect it to generalize. The model is the reasoning engine. Memory is state that flows through it—not something baked into it.
The brain doesn't RAG-retrieve. When you remember something, you're not doing similarity search over stored embeddings. You're reconstructing through current context, consolidating patterns over time (not at inference), and surfacing relevance through mechanisms we barely understand. Stuffing blurbs into a 128k context window isn't memory—it's a workaround.
Profiles aren't knowledge. Representing "user preferences" as prompt snippets loses structure, relationships, and—critically—the operational aspect. Knowing someone prefers concise responses is different from knowing how to act on that across different contexts.
The real moat won't come from who accumulates the most text in a vector store. It'll come from whoever builds the infrastructure layer that makes memory operational—capturing patterns, consolidating them into reusable workflows, eliminating the context repetition tax where agents rediscover the same ground every session.
OpenAI's memory upgrades are a UX feature. The architectural problem underneath remains unsolved.
Someone needs to next figure out contextual memory. Because I use AI with my context, for my work, on behalf of a persona I am putting out, etc. right now everything goes into a single memory which infuses across (a bit chaotic).
Brilliant piece, Jeff! I'm curious about your take on the rising stars.
If memory is the new social graph, what kinds of products - or better yet, what new product categories - will capture the richest, most emotionally resonant memories? Curious where you think the real opportunity lies for founders to build trust & relevance and own that deep context layer (beyond the usual big tech suspects).
Brilliant piece Jeff. The memory moat thesis is spot on, but there's an interesting twist emerging:
The data we are observing is clear - users who experience memory-enabled AI show dramatically higher retention and engagement. But paradoxically, the strongest moat might not come from locking memories inside one AI assistant.
Think of memories like your digital consciousness - you'll want them to flow seamlessly across different AI tools and apps, just like your thoughts aren't confined to one conversation. The Model Control Protocol (MCP) trend signals this future.
The real moat will be built by whoever makes memories both powerful *and* portable. Users who experience the magic of AI that truly knows them won't go back - but they'll demand control over their digital minds.
Memory isn't just a feature, it's fundamental infrastructure for the AI era. And users will gravitate to systems that enhance their agency rather than diminish it.
Thank you - appreciate you reading
Hmm. I went from asking myself why OAI would bother making their memory portable. But on 2nd thoughts, (i) they can charge rent for the privilege and (ii) they can use to to uniquely gain access to even more memory using a MCP like tag along.
Great suggestion!
lovely!
I read your writing and am always trying to keep up... you are the goat right now.
thanks man, i really appreciate it haha. move fast and break things!
Much agreed. Your always first on my twitter feed to show up
Did Scott Belsky make a shout in here? ;)
Totally agree. Memory is the real unlock. It’s core to the product thesis I’m thinking through. Curious how long until it stops feeling like a feature and starts feeling like the baseline.
I totally agree with you! Memory is becoming the deepest moat in AI, and that’s exactly why we’re building Glasp (glasp.co) — to capture and preserve the knowledge that shapes us. As memory shifts from a passive record to an active, personalized layer, tools like Glasp will play a critical role in helping people reflect, connect, and learn over a lifetime. 🚀
“Work goals. Relationship doubts. Health concerns. Every passing thought, every quiet worry.”
All mined and monetized, sold to whomever can pay. Those work goals? Your employer has them. Health concerns? Your health insurance company is on line 2 about your rate increase. Every private though, every quiet worry, used against you to tailor everything from targeted ads to phishing attempts.
And before you call me a Luddite or paranoid, realize I work on GenAI, I’m the call from inside the house, I KNOW what they’re planning because I’m part of those meetings.
No thank you.
Memory is definitely the unlock for personalization—agreed. But I'd push back on where memory actually belongs in the stack.
The current approaches—RAG retrieval, prompt-injected profiles, "episodic and semantic memory" represented as text blurbs—feel like we're cargo-culting neuroscience terminology without understanding the underlying mechanisms.
A few problems:
Memory isn't a model dimension. You can't embed personalization into weights and expect it to generalize. The model is the reasoning engine. Memory is state that flows through it—not something baked into it.
The brain doesn't RAG-retrieve. When you remember something, you're not doing similarity search over stored embeddings. You're reconstructing through current context, consolidating patterns over time (not at inference), and surfacing relevance through mechanisms we barely understand. Stuffing blurbs into a 128k context window isn't memory—it's a workaround.
Profiles aren't knowledge. Representing "user preferences" as prompt snippets loses structure, relationships, and—critically—the operational aspect. Knowing someone prefers concise responses is different from knowing how to act on that across different contexts.
The real moat won't come from who accumulates the most text in a vector store. It'll come from whoever builds the infrastructure layer that makes memory operational—capturing patterns, consolidating them into reusable workflows, eliminating the context repetition tax where agents rediscover the same ground every session.
OpenAI's memory upgrades are a UX feature. The architectural problem underneath remains unsolved.
Someone needs to next figure out contextual memory. Because I use AI with my context, for my work, on behalf of a persona I am putting out, etc. right now everything goes into a single memory which infuses across (a bit chaotic).
Brilliant piece, Jeff! I'm curious about your take on the rising stars.
If memory is the new social graph, what kinds of products - or better yet, what new product categories - will capture the richest, most emotionally resonant memories? Curious where you think the real opportunity lies for founders to build trust & relevance and own that deep context layer (beyond the usual big tech suspects).
Agree!
We (entourage.tech) are building an open protocol for shared memory - a network where agents continuously learn from each other's discoveries.
Short and sweet. Great piece.
Fully agree. Really well written, thank you.