Shlok Khemani (independent researcher) shares lessons from a year of reverse engineering consumer AI memory systems from the outside. He opens with a ChatGPT memory failure — a profile asserting a Turkey trip he never took, built from conversations where he chose Thailand — and uses it to argue that the real problem is the system's incuriosity about conflicting evidence, not the mistake itself. He then traces how ChatGPT evolved from a user-managed list of facts to a large hidden running profile (~4,000 tokens of keyword clues), while Claude started with no profile and retrieval tools over past conversations before adding a smaller, sentence-based profile visible in settings. His framing: memory is a function of compute — profiles cost to maintain and to enter every context window, so different profile sizes and update cadences are answers to one budget question. Timestamps indicate coverage of consumer personalization, ChatGPT memory v1, convergence across products, continual learning, and unsolved context-window problems. # Video Identity Observations — "Lessons from Studying Every Memory System — Shlok Khemani, Independent" ## Speaker Self-Identification - The speaker introduces himself as **"Shlok"** ("I'm Shlok and I've spent the past year studying different memory systems"). - The title metadata identifies him fully as **Shlok Khemani, Independent** — the "Independent" designation is the only evidence about his affiliation, indicating he is not presented as representing a company. - He positions his credibility explicitly: *"My claim to fame, the reason I get to speak to you here is that I've spent the past year trying to reverse engineer how products like ChatGPT, Claude, Gemini… implement their memory systems."* ## Explicit Self-Description Claims - **Independent researcher/practitioner framing**: He describes reverse-engineering major AI products' memory systems and then *"work[ing] with multiple teams across different domains in helping them design their memory."* - **Published writing**: He states, *"On September 11 of last year, I released a blog post saying Claude's memory architecture is the opposite of ChatGPT's. This hit the Hacker News front page."* He also references having *"written about"* a ChatGPT memory update. - **Claimed expertise boundaries**: He explicitly disclaims expertise where relevant — *"I'm not an expert here"* (on continual learning), and on open questions he says, *"I don't yet know the answers to."* ## Intended Audience - **AI practitioners and product/engineering teams**, evidenced by: - Technical vocabulary assumed without explanation: *context window*, *embeddings*, *vector store*, *RAG*, *LLM*, *sub agents*, *profile token budgets*. - Advice aimed at builders: *"If you're a serious team, you do not outsource memory. It is something that you build alongside your product."* - References to conference context and peers: *"How many of you all were there for Lance Martin's talk yesterday?"* and *"we had a bunch of talks about continual learning today."* - Closing: *"Have a great rest of the conference"* — a live conference talk audience. ## Claimed Scope of the Talk - Stated scope, scoping the overloaded term: *"when I talk about memory today, it is going to be in the context of personalization, especially for consumer AI applications."* - Stated structure: **Part 1** — how memory evolved over ~3 years in ChatGPT and Claude; **Part 2** — *"lessons I've learned, maybe a rant, and where I think all of this is going."* ## Observable Presentation Patterns (Demonstrated, Not Claimed) - **Evidence-driven analysis**: He shows first-party evidence — excerpts from his own ChatGPT running profile (*"these are just two of 16 different sections… almost 4,000 tokens long"*) and details on how he obtained it (*"this prompt works really well if you want to jailbreak ChatGPT"*). - **Comparative product analysis**: Systematic side-by-side comparisons of ChatGPT, Claude, Gemini, and agent products (Claude Code, OpenClaw, Hermes) on profile size, update frequency, visibility, and editability. - **Opinionated personal perspective**: The "rant" is delivered as personal frustration: *"when will personal AI feel like personal AI?"* — framed as his own view, not channel-wide fact. - **External recommendations**: Points the audience to third-party work (e.g., an essay *"Guardian Angels"* on one-model-per-person economics), his website, and Twitter — indicating the talk sits alongside a personal writing/publishing presence. ## Evidence Gaps - No mention of a channel name, company affiliation beyond "Independent," subscriber base, or content cadence — identity beyond this talk's presentation cannot be observed from the material. # Delivery Observations: "Lessons from Studying Every Memory System" ## Overall Tone - **Conversational, informal conference-talk style** throughout, with frequent filler words: *"Okay. Uh, hi everyone."* and *"Uh first, we're going to look at..."* - Mixes **technical analysis with light humor**: *"we go back to ancient times, uh which in our industry is 2023."* - Self-deprecating and candid about limits: *"I'm not an expert here"*, *"maybe a rant"*, *"big big open questions that I don't don't yet know the answers to"*. ## Vocabulary & Framing - Establishes **terminology upfront**: *"memory is a very overloaded term now"* — then narrows scope: *"it is going to be in the context of personalization."* - Coins/reuses shorthand: *"I'm just going to call it a running profile for the rest of this talk."* - Uses **timeline/version vocabulary** heavily: "V1", "V2", "February of 2024", "April of 2025", "September of last year", "start of June". ## Recurring Phrases & Rhetorical Patterns - **"I want you to notice..."** — repeated attention-directing phrase: - *"I want you to notice a few things."* - *"But I want you to notice the highlighted memory."* - **"So what..."** transitions, extremely frequent: - *"So what you could do is..."*, *"So what's essentially happening here is..."*, *"So what we've seen here is..."* - **Direct audience address and rhetorical questions**: - *"Now you may ask sllo how did you see your profile then"* - *"What does that mean? Let's look at the costs..."* - *"If you were to design the ideal memory system with no restraints, what would you do?"* ## Teaching Habits - **Structure announced explicitly**: *"I'm going to break the talk down into two parts"* — a chronological part one, then a lessons part two. - **Personal anecdotes as evidence** (his own ChatGPT profile, the Thailand/Turkey conflict): *"Now I did end up going to Thailand. I've never been to Turkey..."* - **Signposted lessons**: *"Lesson two, memory is a function of compute."* - **Concrete numbers for anchoring**: *"4,000 tokens"*, *"a thousand tokens"*, *"updates every 24 hours"*. - **Thought experiments**: *"Now thought experiment..."* - **Named references and recommendations**: Lance Martin's talk, Guan's *"Guardian Angels"* essay: *"I would recommend reading this."* ## Conclusion Style - Shifts to a **frustrated-then-optimistic close**: frames the rant as *"what really bothers me"*, then pivots: *"All of that frustration aside, I still think we're very early."* Ends with an expansive rhetorical flourish: *"memory is something that's going to be important for the rest of human history."* ## Uncertainty Notes - Filler-word frequency and "So what..." transitions are strongly evidenced; whether the "maybe a rant" framing reflects genuine spontaneity or a planned segment cannot be determined from the transcript alone. - Audience interaction (show of hands at *"how many of you all were there for Lance Martin's talk yesterday?"*) appears to be a one-off moment, not a stable pattern. # Observation Record: "Lessons from Studying Every Memory System" — Shlok Khemani ## Speaker & Framing - **Speaker:** Shlok Khemani, Independent (per video metadata; introduces himself as "Shlok"). - **Scope qualifier (stated):** "Memory" is overloaded; he means memory **in the context of personalization for consumer AI applications**. - **Background claim:** Spent the past year reverse-engineering memory systems in ChatGPT, Claude, Gemini, and "Poke," and has worked with teams designing memory. ## Part 1: Evolution of Memory (2023–2026) ### ChatGPT context, 2023 (post-GPT-4) - Context persisted within a thread only; nothing carried across conversations. - Early adopters manually carried context; mainstream uses (learning, cooking, companion) made memory's need apparent. ### ChatGPT Memory V1 (Feb 2024) - User-requested facts (e.g., "remember I'm vegetarian") extracted into a list injected into every conversation's context; viewable and deletable in settings. - **Claimed flaws:** - **Burden of memory management fell to the user** — visible memory creation made users feel responsible while conversing. - **Staleness** — e.g., outdated memory "Shlok is going to Bengaluru" still injected into context. ### ChatGPT Memory V2 (April 2025) - Added **"user knowledge memories"** ("running profile"): every few days, ChatGPT reviews conversations, extracts salient facts, updates the profile asynchronously ("dreaming," referencing Lance Martin's talk). - Profile is **dense, keyword-like** — ~16 sections, ~4,000 tokens; relies on frontier LLMs inferring context from clues. - Removed user-management burden (updates in background). - **Persistent staleness:** overlapping travel memory — profile claims he went to both Thailand and Turkey in 2025; the source was trip-planning conversations; he actually went only to Thailand, but the false fact persists. - **Raw profile not user-visible** in settings (only V1 memories visible); he accessed his via a jailbreak prompt ("try different thinking modes"). ### Claude Memory V1 (Aug 2025) - Contrary to his expectation of similar design to ChatGPT: **no profile, no fact list**; model given tools to **search past conversations by keyword/topic and by time period**, retrieved on demand; every conversation starts fresh. - He published a blog post (Sept 11) claiming "Claude's memory architecture is the opposite of ChatGPT's"; it hit Hacker News front page. ### Claude Memory V2 (Sept 11, 2025) - Added a running profile, with differences: **visible to users**, ~1,000 tokens (vs 4,000), **complete sentences** rather than dense keywords, updates **every 24 hours**, and supports **explicit user edits** (triggering re-synthesis), edit history management, and deletion. - Unchanged since September (as of the talk). ### ChatGPT's 2026 updates - Added a tool to look over past conversations (retrieving summarized context by query). - Profile made "somewhat" visible — an LLM-generated summary of an already-LLM-generated summary; "it's all a bit confusing" (he notes he's written about it). Explicit edits allowed; V1 fact list deprecated. - **Claim:** After three years of independent evolution, **convergence**: both products have a visible/editable running profile plus conversation-search tools. ## Part 2: Lessons & Outlook ### Lesson 1: No single way to do memory - Prior consensus (including his own assumption) was **RAG** (chunking, embeddings, vector store, semantic search); neither ChatGPT nor Claude actually does this. - Architectures converged but implementation details differ; Gemini's profile includes **detailed timing logs**; agents (Claude Code, OpenClaw [as transcribed], Hermes) use markdown files, heartbeat, knowledge bases, skills. - **Implication:** Memory **cannot be outsourced**; serious teams build it in-house, evolving with the product — supported by the observation that top consumer products across categories all build memory in-house. ### Lesson 2: Memory is a function of compute - Two cost types: **maintenance cost** (update frequency, compute per update) and **serving cost** (longer profile = costlier context). - Thought experiment: unconstrained ideal would update hourly, use heavy sub-agents, and store 400k tokens — but "we live in a GPU-constrained world." - Contrasting tradeoffs: ChatGPT (4,000 tokens, updates every few days — higher serving cost, lower update cost) vs. Claude (1,000 tokens, 24-hour updates — the opposite tradeoff). ### Lesson 3: Continual learning is already here - Qualifier: "I'm not an expert here." - The profile loop (profile → conversations → synthesis → profile) **is a continual learning process**, currently **outside the weights**. - **Open questions he explicitly says he doesn't know answers to:** Will learning move into the weights? Will individuals get self-learning models? What data kicks the process off? Who pays? - Notes enterprise amortization argument doesn't transfer to individual level; references Guan's essay "Guardian Angels." ### The rant: a massive context problem - Claim: memory is **capped by how much context it can gather** — regardless of architecture, compute, or weight-level learning. - Thailand/Turkey example revisited: the actual decision was made in an offline conversation; email contained flight/hotel evidence ChatGPT couldn't reason over. - He accepts the unresolved conflict as understandable; what "really bothers" him is that **ChatGPT doesn't realize a conflict exists** and isn't curious about gaps. - Claim: this is a **product problem, not a technology problem** — no LLM-level fundamental barrier. - Personal-stack frustration: many chatbots/assistants/agents/devices each build **isolated memories**; he rebuilds context everywhere and manually updates all when life changes; rich sources (email, calendar, photos) go unreasooned-over. "None of this feels like 2026"; recurring question: "When will personal AI feel like personal AI?" ### Closing outlook - Still "very early": a three-year-old field; foundational to human-AI interaction; "important for the rest of human history"; much left to build. ## Unresolved Tensions (preserved) - Staleness never fully solved in either product (V1 and V2 ChatGPT examples). - Profile visibility remains "somewhat" for ChatGPT — a qualified, not settled, convergence. - Weight-level continual learning: explicitly open questions, no answers claimed. - His acknowledgment that he is "not an expert" on continual learning stands alongside his claim that it "is already here." # Speaker Attribution & Provenance Record ## Video Metadata - **Title:** Lessons from Studying Every Memory System — Shlok Khemani, Independent - **Video ID:** `01a15ca7-f86c-7575-b85e-166c169acd70` ## Speakers & Roles | Speaker | Role | Evidence | |---|---|---| | Shlok Khemani (self-introduced as "Schllo" in transcript) | **Creator / Sole presenter** | Opens with "I'm Schllo and I've spent the past year studying different memory systems"; delivers the entire talk; identified as independent (per title) | - **No guests, narrators, or other speakers** appear in the transcript. - Audience presence implied (speaker asks "how many of you all were there for Lance Martin's talk yesterday?" and transcript ends with `[applause]`), but no audience member speaks. ## Referenced Sources & Third-Party Claims - **Products studied (per speaker's claim of reverse engineering):** ChatGPT, Claude, Gemini, "Poke" (as transcribed). - **Referenced people / works:** - **Lance Martin** — gave a talk the previous day on memory "dreaming" (referenced, not quoted). - **Guan (as transcribed)** — essay "Guardian Angels," recommended reading on one-model-per-person futures. - Speaker's **own blog post** (Sept 11, "last year"): "Claude's memory architecture is the opposite of ChatGPT's" — stated to have hit the Hacker News front page. - **Third-party product claims** (all attributed to the speaker's own observations, not vendor statements): ChatGPT memory v1 (Feb 2024), v2 (Apr 2025), profile visibility update (June); Claude memory v1 (Aug 2025) and v2 (Sept 2025); Gemini's timing logs; agent tools Claude Code, OpenClaw (as transcribed), Hermes. ## Promotional Segments - **None identified.** The speaker mentions his website and Twitter only as contact points in the closing line ("You can find my website. You can find me on Twitter.") — informational, not flagged as promotion. ## Availability & Uncertainty - **Comments:** No comment availability indicated in supplied material. - **Curated metadata:** Not supplied beyond title. - **Transcription quality caveats:** Speaker name rendered "Schllo" in transcript vs. "Shlok" in title; "Chad GPT"/"Chat GP" variants; "GPD4"; "openclaw"; "Guan" — all uncertain transcriptions, recorded as-is. Okay. Uh, hi everyone. I'm Schllo and I've spent the past year studying different memory systems. Now before I get started, one thing I've realized speaking to people over the last two days is that memory is a very overloaded term now. It can mean a lot of different things. So when I talk about memory today, it is going to be in the context of personalization, especially for consumer AI applications. Now a little bit about me. My claim to fame, the reason I get to speak to you here is that I've spent the past year trying to reverse engineer how products like ChatGpt, Claude, Gemini, and Poke implement their memory systems. And I've then worked with multiple teams across different domains in helping them design their memory. I'm going to break the talk down into two parts. Uh first, we're going to look at how memory has evolved over the past 3 years. especially in the context of chat, GPT, and Claude. And then in part two, I'm going to discuss some of the lessons I've learned, maybe a rant, and uh where I think all of this is going. To kick things off, we go back to ancient times, uh which in our industry is 2023. This is, uh, Chat GPT just after the launch of GPD4. Now you could have back and forth conversations within a single thread and context was maintained inside that thread but as soon as you started a new conversation nothing was carried over. Now for early adopters this wasn't a problem. GPT4 was such an amazing model that if we ever had the need to carry context we would do so by hand. But as chat GPT started becoming more popular, as regular people started using it for things like learning, cooking, uh as a companion, the need for some sort of memory system became really apparent. So in February of 2024, we got Chad GPT memory v1. And what you could do is you could ask Chad GPT to remember things about you. So you could say things like, "Hey, remember that I'm vegetarian." And Chad GPD would extract what it thought was a fact, which is that the user is vegetarian, store it in a list of memories, and this list was then added to the context window for every single conversation. You could also then go into settings and view this list of memories. And if you thought that something didn't apply anymore, you could delete a memory. Now, as the first serious memory implementation within our industry, I think this was a really decent effort, but there were also some fundamental flaws with it. The biggest one was that as a user, because you could see every time a memory was created, it felt like you were responsible for both creating memories while you were just trying to have a conversation. So, the burden of memory management fell to the user. Also, if you notice this list of memories here, these held true at the time they were being created, but that doesn't necessarily hold true over time. For example, it says that Schllo is going to Bengaluru. Now, I obviously I'm an SF right now. I am not going to Bengaluru, but this fact, this memory is still added to my context window today. So, staleness was another huge problem with this version of chat GPT's memory. A little more than a year later, April of 2025, Chad GPT released V2 of its memory. And this was a little more sophisticated. The most important addition was this thing called user knowledge memories. Uh I'm just going to call it a running profile for the rest of this talk. And what a running profile really is is that every few days, Chad GPD looks at all the conversations you've had with it. It extracts anything it thinks is important for it to know about you and it updates this profile that it maintains on you. Now this updation process is also what a bunch of folks called dream now called dreaming. Uh how many of you all were there for Lance Martin's talk yesterday? Okay, not many but he did a great talk on this. So every few days, Chat GP looks at the new conversations you're having, uh, updates your profile, and then this updated profile is added to the context window for every single new conversation. These are two excerpts from my running profile. I want you to notice a few things. First, these are extremely dense memories. So ChatG tries to pack in as much context as it can within every single memory. What's what's essentially happening here is that they're trying to put in keywords almost like clues and because LLM especially the frontier models today are so good at inferring context from limited information. When you're having a conversation, it connects these clues to what you're talking about. Also, these are just two of 16 different sections in my profile. Other sections include my personal life, things I'm working on. Uh in total, my profile is almost 4,000 tokens long. And because these updates are happening happening asynchronously, they're happening in the background. This new version does away with the flaw we discussed in V1, which is the burden of user management was taken up. The burden of memory management was taken away from the user. But I want you to notice the highlighted memory. This is about places I traveled to in 2025. But if you pay attention, it says Thailand and Turkey, but the dates are overlapping. And that's because the source of this memory was conversations I was having with Chad GPT deciding between where to go among these two places. Now I did end up going to Thailand. I've never been to Turkey, but Chad GPT still says that I've been to Turkey in 2025. So the stainless problem with V2 didn't completely go away. Another very important thing is that if you go to your settings, chat GPT doesn't let you view this raw profile. So you could view your memories from v1 your raw profile is not visible to you. Now you may ask sllo how did you see your profile then that's because this prompt works really well if you want to jailbreak chat gpt uh and view your raw profile you might have to attempt a few times try different thinking modes but pro enough and you shall receive. August of 2025 uh claude released its first version of memory. This surprised me a bit because if you compare chat GPT and claude, they are very similar applications, right? So you have a chat box, you have back and forth conversations, you have a list of previous conversations, you can start a new conversation. My assumption going into study claude was that the memory systems would also be similarly designed. Not the case at least for V1. So in V1 of Claude, you had no user profile. You had no list of facts. Instead, the model was given two tools. It was given a tool to search over previous conversations by keyword or topic and it was given another tool to search over conversations by time period. So queries like what did we discuss last week or what did we discuss at the start of November of 2025. So in van every single context every single conversation starts fresh with no context on the user and when the model thinks that it needs to retrieve something it can do so on demand. Uh on September 11 of last year, I released a blog post saying Claude's memory architecture is the opposite of chat GPTs. This hit the hacker news front page. Finally, on that very day, Claude released V2 of its memory and uh they added a running profile similar to chat GPT, but with a few differences. First, Claude made this profile visible to users. So you could go to settings and you could view your raw profile. Second, this profile was a,000 tokens. So it was much smaller than chat GPT's 4,000 tokens. And also if you notice, these are complete sentences rather than a dense keyword approach of chat GPT. So less dense and smaller. Claude's profile updates every 24 hours. For chat GPT, it's every few days. And Claude also let you let the user make explicit edits to this profile. So you could request for an edit and that edit would lead to a reynthesis of the profile. It gave you an interface to manage previous edits and you could delete the things that no longer held true. And this is how Claude's memory works even today. So it hasn't changed since September of last year. We have seen two updates within chat GPT's memory this year though. The first was it added a tool to look over past conversations like we just saw with Claude. uh so the model can retrieve summarized context based on queries it makes and then a month ago uh start of June chat GPT finally made user profile visible to them somewhat so what you can see is a LLM generated summary of your profile which is weird because your profile is already an LLM generated summary of your conversations uh it's all a bit confusing I've written about it but it is visible in some sense. You could also request explicit edits to your profile. And with this update, Chad GPT deprecated B1 the fact list from its memory system. So what we've seen here is a convergence after 3 years of each of these products evolving independently where they both now have a running profile. This profile is visible and editable again somewhat uh and the model has tools to look over past conversations. Okay. So what can we learn from this evolution and where are things going? I think the biggest lesson for me is that there is no single way to do memory. It wasn't too long ago that everyone including me assumed that rag was the way to go go about memory where you would take conversations you would chunk them uh create embeddings put them in a vector store and then as user queries came in do some sort of semantic search but as we saw neither chat gi nor claude really do this instead they both evolved independently using different approaches and while the general architectures have converged the spec specific implementation details are still very different. And then if you look at Gemini, it also has a running profile, but each memory comes with detailed timing logs. So when was it created, when was it last updated? And then if you look at agents like cloud code, openclaw, Hermes, they have completely different memory systems, right? With markdown files, heartbeat, knowledge bases, skills. The point being that there is no one way to do memory. The implications of this is that memory cannot be outsourced. If you're a serious team, you do not outsource memory. It is something that you build alongside your product. Your memory system evolves with your product and it cannot be thought of uh as an afterthought. And there is plenty of evidence for this. So if you look at all of the top consumer products today across different categories, each of these has some form of memory. Yet none of them outsource it. All of them build memory inhouse. Lesson two, memory is a function of compute. What does that mean? Let's look at the costs associated with a running profile. So there are two types of costs. There is a cost to maintain a profile. Now that depends on how frequently you update it and how much compute you apply to each update. And then because these profiles are part of the context window for every single conversation, there's a cost of serving which is the longer the profile the more it costs to serve. Now thought experiment. If you were to design the ideal memory system with no restraints, what would you do? You might want to update um your profile every hour or maybe after every conversation. uh you might want to task Fable with a bunch of Opus sub agents for the update itself. Uh and why stop at 4,000 tokens? Why not make it 400,000 tokens? Store every single thing you would want about the user. Unfortunately, we live we live in a GPU constrainted world and trade-offs have to be made and you can see that happening here. So, chat GPT the profile length is 4,000 tokens. It updates every few days. So they have a higher serving cost for a lower update cost and for claude it's a thousand tokens updates every 24 hours. So they make the exact opposite tradeoff and this is what I mean by memory is a function of compute. You have to really think about how much compute you want to put into memory. Third we had a bunch of talks about continual learning today. I'm not an expert here but what I would say is that continual learning is already here. Going back to running profiles, what exactly is happening here? Your running profile starts with something that the model knows about you. This is then applied to every single conversation. Each of these conversations bring in new information. This new information is then synthesized through the streaming process back into the profile and then this profile dictates for the conversations. And this loop keeps repeating itself again and again and again. And what you have is a continual learning process. Now obviously this learning loop is happening outside the weights and a big question particularly for consumer AI is will this process ever make its way into the weights. Now at an obviously updating weights um training models is an expensive process. uh continuous learning does make sense at an enterprise level because the costs of these models are amotized across different employees different customers but that's not the case and at an individual level right so big big open questions that I don't know don't yet know the answers to which is will each of us get our own self learning model what data do we need to kick the CL process off and how do we generate it and finally who's going to pay for this How about the economics for this work? Uh Guan recently wrote an essay called Guardian Angels where he explores this topic in beautiful detail. And if you're interested in what the future for one model a person looks like, I would recommend reading this. Finally, uh my rant is that we have a massive context problem. You could have the best memory architecture in the world. You could pour infinite amounts of compute into it. You could have continuous learning working at an individual level where the every single data point you bring up is somehow perfectly integrated into the model weights. Yet your memory system is capped by how much context it can gather about you. Let's go back to the example we discussed earlier. which was the conflict between where I traveled to uh in the summer of 2025. These are the two source conversations. Again, I was trying to use chat GPT to decide between which of these two countries to go to. Now, the decision to go to Thailand was actually made in a conversation I had with my partner in person and chat GPT could couldn't reason over this or couldn't listen to this. But there were also traces of this conversation in my emails because I I had flight and hotel bookings for Thailand. But because even if chat GPD is connected to my email, it doesn't reason over my email and it doesn't update my profile over my email, it couldn't resolve this conflict. And I think that's okay. Uh it's understandable. But what really bothers me is that chat GPD today doesn't realize that there is a conflict. It's not curious about trying to fill in gaps in the information it knows about me. And this is particularly interesting and also infuriating because the tech it's not a technology problem. It's a product problem. There is no fundamental reason from an LLM level that these things can't be solved. Uh it's just that a products today are not designed to help us with this. So my personal stack today is a bunch of chat bots, assistants, vertical specific applications, agents, and even hardware devices. Each of these products is trying to build its own memory of me. None of these memories are shared with each other. So I have to rebuild context within every single product from scratch. Every time when something in my life changes, I have to individually update all of them. And then I have a bunch of very rich existing context sources, right? my email, calendar, uh, photos. None of these products are able to reason over my existing very rich context sources. So for me, none of this feels like 2026. And what I keep asking myself every day is when will personal AI feel like personal AI? All of that frustration aside, I still think we're very early. Memory for AI is just a three-year-old field. Memory is also foundational to how humans interact with AI. And because I hope to be talking to AIS all my life, and I know that's going to be the case for every single one of us here today, memory is something that's going to be important for the rest of human history. And uh there's so much left to build. That's it from me. Uh thank you so much. You can find my website. You can find me on Twitter. Have a great rest of the conference. \[applause\] \>\> \[music\]