Dylan Couzon, developer relations engineer at Qdrant, argues that while frontier-class AI models can now run on hardware you own, the memory that personalizes an assistant typically remains hosted in external clouds. He breaks memory down into write, retrieve, and forget operations, models the stack as CPU = model, RAM = context, and disk = memory, and explains why retrieval outperforms storing everything in the prompt. The talk includes a live fully offline demo of a drone building searchable memory with sub-millisecond queries in a 15 MB footprint using Qdrant Edge, plus a discussion of opt-in shared cloud-synced 'hive mind' memory. Recorded at the AI Engineer World's Fair 2026 in San Francisco. # Channel Presentation Observations ## Source Scope - **Video ID:** `01a15ca7-2254-7a99-a60e-6cec0701a68f` - **Title:** "Stop Renting Your AI's Memory — Dylan Couzon, Qdrant" - **Curated metadata:** None supplied. Identity context comes only from the title and transcript. - **Format:** A single conference talk. No timestamps are available, so none are cited. ## Explicit Self-Description These statements come from the speaker, so they are creator claims. - **Role:** Introduces himself as a developer relations engineer. > "I am a developer relations engineer at Quadrant" - **Affiliation:** Names Qdrant as his employer. The transcript spells it "Quadrant," likely a transcription artifact, since the title reads "Qdrant." - **Talk framing:** Gives the title as "the frontier is coming home" and states the thesis as "who it belongs to." - **Stated scope of his own product:** > "I work in a vector search engine that embeds directly into your app." - **Self-positioning:** Presents Qdrant's approach as one option, not the only one. > "this is the piece that I work on, and it's just one way to do it." He also says he won't argue "which memory architecture wins." ## Observable Presentation ### Intended Audience The transcript shows the speaker addressing a live conference room. - Opens with conference-directed remarks: "I hope everyone is having a great conference so far." - Uses inclusive address throughout: "everyone here," "everyone on Earth" (in context of AI infrastructure). - Closes with an audience-directed call to action: > "tonight, don't go home and just try your tool." The content assumes technical familiarity: embeddings, HNSW, quantization, vector search, agents, and token costs appear without definition. ### Claimed Scope The talk claims a broad thematic territory: - Ownership of frontier-class AI (local compute, open weights) - Persistent, personal memory as the missing layer in AI systems - Data sovereignty and opt-in sharing of memory It makes several factual assertions without sources in the transcript, including a government order affecting model access, subscription subsidy claims, and hardware performance claims. These are presented as the speaker's statements. ### Demonstrated Patterns These are observable in the material itself, not just asserted. - **Live demonstration:** A drone video is processed with YOLO object detection, and detections are embedded into a local vector store. > "this is actually running live right now, fully offline on my laptop." - **Quantified demo outputs:** The speaker reports figures from the demo, including "92 different objects," "over three 300 vectors," and a footprint of "only 15 megabytes." These are specific numbers spoken during the demo, not independently verified here. - **Product-embedded content:** The product is shown inside the argument. The demo, sync capability, and performance claims all point to the speaker's employer's engine. - **Rhetorical structure:** Recurring contrast between "rented" and "owned," and between "stranger" and "private mind," frames each section. - **Narrative arc:** The talk moves from problem (rented AI), to diagnosis (memory as continuity), to solution (local vector memory), to a call to action. ## Creator Claims vs. Demonstrated Evidence | Claim | Type | Evidence in Material | |---|---|---| | Frontier-class models can run on consumer hardware under $2,500 | Stated claim | Spoken only; no demonstration shown | | Local memory can hold a million entries under 1 GB with quantization | Stated claim | Spoken only; the drone demo shows 15 MB for about 300 vectors | | Semantic retrieval over local memory runs in under 1 ms | Demonstrated (as shown in demo) | Speaker reports timing during the live coffee-table query | | Memory, not inference, is the primary lock-in risk | Argued position | Argument, not evidence | | Sharing should be opt-in | Stated principle | "sharing should always be opt-in, never the default" | ## Limits of This Observation - Only one video is available. Nothing here supports conclusions about the channel as a whole. - The transcript contains likely ASR errors (for example, "Quadrant," "Cozan," and garbled phrases near the close), so exact wording may differ from the original audio. - Factual claims about external events, pricing, and model availability are unverified within this material. - No private motives or business intentions are inferred from the employer relationship. # Delivery Observations: "Stop Renting Your AI's Memory" (Dylan Couzon, Qdrant) **Evidence base:** Title and transcript only. No audio, video, or curated metadata was provided, so this document covers **word choice, sentence construction, and structure** only. It does not cover pace, pitch, gesture, or vocal emphasis. The transcript appears to be auto-generated, so some repetitions (for example, "the the") may be transcription artifacts rather than spoken delivery. --- ## 1. Verbal Fillers and Conversational Markers - **"you know"** recurs throughout as a discourse marker, often mid-argument: > "And you know, there's an obvious pushback." > "So, you know, today I'm not going to argue which memory architecture wins or is the best one." - **"Um"** appears at the opening and in the demo section: > "Um so my name is Dylan Cozan." - **Repeated words** appear in several places, possibly from transcription: > "But the the assistant you're using right now forgets..." > "What make What makes it yours is everything it knows about you." **Uncertainty:** The fillers are consistent, but whether the repeated words are spoken stutters or caption errors cannot be determined from text alone. --- ## 2. Hedging and Qualified Assertions The speaker often softens claims with "I think," "kind of," or "basically," and sometimes names the limits of a claim directly. - **"I think"** frames opinions as perspective: > "I think when is kind of the boring question. I think the real one is who it belongs to." - **Explicit caveats on capability:** > "Not the absolute frontier, I'm going to be honest, but shockingly close." - **Scope limits:** > "So, you know, today I'm not going to argue which memory architecture wins or is the best one. That's definitely a different talk." Hedges appear mostly around opinions and capability comparisons. Factual statements, such as storage sizes and latency figures, are stated without qualification. --- ## 3. Recurring Rhetorical Structures ### 3a. Negation-then-correction ("not X, Y") This is the most frequent contrastive pattern in the transcript: > "Not with a bigger prompt, with actual memory." > "It's not a bigger prompt. It's a system with three verbs: write, retrieve, and forget." > "It's not a smarter you. It's a you that doesn't forget." > "That's not a smaller super intelligence, that's the only one worth wanting." ### 3b. Rhetorical questions as section openers Questions often introduce a new section, followed by a direct answer: > "What would you build if your assistant actually remembered everything you ever taught it?" > "So, what actually is memory?" > "Why not just throw everything in a big folder of markdown files?" > "Where did I leave my badge? What was her daughter's name again?" ### 3c. Triads and lists of three Three-part lists appear in both claims and examples: > "The compute is rented. The model is rented. The harness is rented." > "write, retrieve, and forget" > "Every preference, every correction, every dead end for years." > "Mom's day, dad's day, your day searchable together." ### 3d. Repetition of a short phrase for emphasis > "Owning inference gives you autonomy... But owning memory gives you continuity." > "Owning" and "rented" recur across sections as the central contrast. --- ## 4. Signposting and Structure The speaker frequently announces transitions with "So," "And," or "Here's": > "So, let's look at what renting actually costs today." > "Here's the deeper issue." > "And one last idea because owning memory doesn't doesn't mean keeping it to yourself." > "And this is where I'm going to leave you." The talk follows a visible arc: framing the problem, diagnosing rented AI, explaining memory, demonstrating it, extending it, and closing with a call to action. Each section is marked with a transition phrase. --- ## 5. Analogies and Named Metaphors Several metaphors are introduced and then reused: | Metaphor | Introduced | Reused | |---|---|---| | **"brilliant stranger"** | "a model on its own is just a brilliant stranger" | "the stranger just disappears"; "Take the stranger home and give it a memory." | | **"second mind"** | "That's not a chatbot any anymore. That's a second mind." | "a private mind that never forgets you" | | **"geniuses with no long-term memory"** | "They're just geniuses with no long-term memory." | Not repeated | | **CPU / RAM / disk** | Attributed to Karpathy: "the model is the CPU, the memory the context window is the RAM, and the memory is the disk." | "the disk is in someone else's data center" | | **"frontier"** | Title: "the frontier is coming home" | "The frontier on your desk"; "frontier models" | The **"stranger"** and **"frontier"** motifs are the most sustained. Attribution of the CPU/RAM/disk analogy to Karpathy is explicit in the transcript. --- ## 6. Direct Address and Audience Positioning The speaker uses second-person address and inclusive first-person plural: > "Hi everyone. Thank you for joining me today." > "every everyone here is going to watch super intelligence show up" > "tonight, don't go home and just try your tool." The closing section shifts to imperatives: > "So tonight, don't go home and just try your tool. Take the stranger home and give it a memory." --- ## 7. Demonstration Narration The live demo section uses a distinct register: - **Present-tense description** of what is happening on screen: > "this is actually running live right now, fully offline on my laptop." - **Precise numeric claims**, stated plainly: > "we have already recognized 92 different objects, and we have over three 300 vectors." > "the entire memory and Quadrant engine footprint is only 15 megabytes." - **Latency claim** with a specific unit: > "in less than 1 millisecond, I was able to pull up every single coffee table" The transcript reads "over three 300 vectors," which looks like an error for "over 300 vectors." This is a caption artifact, not a choice the speaker made. --- ## 8. Sentence Rhythm - **Short declaratives** dominate key claims: > "Weights you own never change unless you change them." - **Long run-on sentences** with multiple clauses appear in explanations: > "Retrieval gives you something a prompt cannot. It is control. You can filter by topic. You can decay by recency and frequency." - **Stacked short sentences** are used for emphasis at section ends: > "Because you were the only one who ever trained it." --- ## 9. Closing Pattern The conclusion combines a reframing claim with an imperative and a vision statement: > "That part is basically settled. What's not settled is who it belongs to, and that's the fight worth having right now." > "Then picture that memory 5 years from now. Not a chatbot, a private mind that never forgets you." --- ## 10. Stable Patterns vs. Isolated Moments **Repeated across the transcript (stable):** - Negation-then-correction contrasts - "you know" as a filler - "I think" and "kind of" as hedges - Triads - Explicit transition phrases - Recurring "stranger" and "rented/owning" motifs **Appears once or twice (isolated):** - Specific demo numbers (92 objects, 15 MB, 1 ms) - The CPU/RAM/disk analogy - The family "hive mind" example - The "smart glasses" reference - The explicit opt-in statement about sharing --- ## 11. Limits of This Analysis - **No audio or video**: Pace, emphasis, pausing, and gesture cannot be assessed. - **Probable caption errors**: Several garbled passages ("Trick Take," "over three 300 vectors," "What make What makes") may not reflect what was said. - **Single speaker, single talk**: Patterns are observed within this transcript only and cannot be generalized to the speaker's other work. - **No metadata**: Audience, event context, and intended format were not provided. # Position Record: "Stop Renting Your AI's Memory" **Speaker:** Dylan Couzon (spelled "Cozan" in the transcript), introduced as a developer relations engineer at Qdrant (transcribed as "Quadrant") **Source basis:** Transcript only. The title and speaker name come from the supplied metadata. No external sources were cited in the talk. **Nature of record:** Positions the speaker expressed. Private beliefs and motivations are not inferred. --- ## 1. Core Thesis The speaker's central claim is that the question of advanced AI is less *when* it arrives than *who it belongs to*. > "I think when is kind of the boring question. I think the real one is who it belongs to." He frames two futures: - **Rented:** Intelligence lives "in someone else's data center, and you pay by the token." - **Owned:** Intelligence "learns from your life, your context stays private, and nobody can switch it off, throttle it, or take it away from you." He argues that the **model weights** can now be owned locally, but the **memory that makes the model personal** remains rented. Ownership of memory, which he calls "continuity," is presented as the missing piece. --- ## 2. Claims by Theme ### 2.1 The Frontier Is Reachable Locally - A machine under **$2,500** today can run "basically last year's frontier." - Qualified immediately: "Not the absolute frontier, I'm going to be honest, but shockingly close." - Open-weights models are "closing on fast," with the gap "shrinking every quarter." - The speaker says this achievement "doesn't quite land" because **a model alone is "just a brilliant stranger."** ### 2.2 Costs and Risks of Renting Inference Three stated risks, each presented as a reason to own the model: 1. **Access revocation.** A government order reportedly "pulled access to two flagship models, Fable and Mythos 5, for every single customer." *(No source given; the model names are reproduced as transcribed.)* 2. **Version churn and throttling.** Providers "retire versions and they throttle what's left when the economics gets tight." Consequently, "the model you rely on can quietly get worse or vanish." 3. **Cost and subsidy.** Agentic workloads consume far more tokens than chat. The speaker claims "some developers burn through over $5,000 a month of compute on a $200 plan," and calls this "a subsidy" that "will ultimately disappear." He contrasts these with local weights: "Nobody can reach into your machine and flip those off... Weights you own never change unless you change them." ### 2.3 Memory as the Real Product - **Owning inference gives autonomy; owning memory gives continuity.** - The speaker asserts that every frontier lab "shipped a memory feature" last year, "not because the models got smarter, but because they don't carry anything between sessions." - He describes continuity as "exactly what every major lab is trying to package up and sell back to you." - He claims that a system that "starts simple and keeps learning you will always feel smarter than one that starts brilliant but forget[s] everything about you." ### 2.4 What Memory Is The speaker defines memory as a system with **three operations: write, retrieve, and forget.** He compares these to human memory. **Reasoning against "a big folder of markdown files":** - Retrieval "beats dumping everything and praying the model finds it." - Retrieval provides **control**: filtering by topic, decay by recency and frequency, and relevance that shifts over time. **Technical readiness claims:** - HNSW indexing dates to 2016. - Local embeddings have run since 2019, "long before any model could really use them at scale." - His conclusion: "The memory side of superintelligence was never the hard part. We were just pointing it at ourselves." ### 2.5 The Memory Hierarchy (Karpathy Analogy) The speaker attributes to Karpathy this framing: | Component | Analogy | Property stated | |---|---|---| | Model | CPU | Computes | | Context window | RAM | Fast, but "forgets the moment the session ends" | | Persistent store | Disk | "Remembers you across every conversation" | His position: "the disk is in someone else's data center" for almost every AI system today, and "the disk, the parts that holds the actual record of you, should be yours fully and forever." ### 2.6 Architecture-Neutral Stance The speaker explicitly declines to argue which memory architecture is best: > "I'm not going to argue which memory architecture wins or is the best one. That's definitely a different talk." He presents his own work as "just one way to do it": a local, embedded vector store, with quantization allowing "a million memories" in "less than a gigabyte," small enough for "a phone or even a Raspberry Pi." ### 2.7 Portability and Persistence - Each app gets an isolated store, and "that folder is the thing that actually follows you." - Stated condition: portability holds "if you keep the same embedding model." Swapping the reasoning model or hardware "none of it matters. It all comes with you." --- ## 3. Live Demo (Drone Object Memory) The speaker ran a demonstration that he described as live and fully offline: - **Setup:** Drone video footage processed with YOLO (object detection). Detected labels were embedded into a live memory store. - **Reported figures:** - 92 distinct object categories recognized - Over 300 vectors stored - Total engine footprint of about **15 MB** - Coffee-table query returned all matches in **under 1 ms**, including image, first-seen and last-seen timestamps, definition, and seen count - **Scope qualification:** Processing was "run and processed locally. There's no network attachment at all." Cloud sync was described as a capability but was **not demonstrated**. The speaker's note that the drone was "first time out in the world" and "doesn't know anything" frames the demo as memory accumulation from zero. --- ## 4. Extrapolations Stated by the Speaker The speaker extends the drone example to other domains. These are **forward-looking proposals**, not demonstrated results: - **Chatbots, agents, and code assistants** that accumulate corrections, preferences, and dead ends. Claimed effect after "a few weeks": "the stranger just disappears." - **Whole-day personal memory:** "Where did I leave my badge?" and similar queries framed as retrieval. - **Wearables:** He states that smart glasses are "already shipping today," and that "that's not science fiction." - **Family and team hive minds:** Separate personal stores that sync into a shared, family-owned memory. --- ## 5. Claims About Existing Data Collection > "Frontier models are already building an index of you today. They have your relationships. They have your habits. They have your whole inner life. That index is coming whether you opt in or not." This is stated as fact in the talk. No evidence is presented in the transcript. The speaker's framing is that the only open question is "whether you own it or if you're renting access to yourself." --- ## 6. Qualifications and Tensions These should remain visible alongside the claims above. - **Frontier gap:** The speaker concedes local hardware is "not the absolute frontier," yet argues the frontier is "on your desk." - **Local-first vs. cloud sync:** The thesis is that memory should be owned and "start local," but the product includes out-of-the-box cloud sync. The speaker frames sync as opt-in: "sharing should always be opt-in, never the default." - **Vendor affiliation:** The speaker works at Qdrant, and his demonstrated technology is Qdrant's engine. He also says he is "not going to argue which memory architecture wins." The transcript does not contain a comparative benchmark against alternatives. The "same raw score as Quadrant in the cloud" claim is asserted without accompanying data in the talk. - **Sharing risk acknowledged:** "A memory can be shared with consent, or it can be extracted without it." He identifies this as the reason to start local, but does not elaborate on mitigations beyond opt-in defaults. - **Unsourced claims:** The government order, the $5,000/$200 usage figure, and the "every frontier lab shipped a memory feature" claim are presented without citations in the transcript. - **Speculative timeline:** "Somewhere in the next few years" and "super intelligence show up over the next few years" are predictive. He also states "That part is basically settled," referring to the arrival of superintelligence, while saying who owns it is "not settled." --- ## 7. Unresolved Questions Raised (Not Resolved in Talk) - Who holds the memory when a device is lost, shared, or sold? (Not addressed.) - How forgetting or decay policies are governed for personal data. (Mechanism described; governance not discussed.) - Whether local-first memory remains practical for models outperforming a $2,500 machine. (Acknowledged as a gap.) - How the "hive mind" consent model handles members who did not opt in. (Only the family example is given.) --- ## 8. Closing Position The speaker's concluding argument: > "Not a chatbot, a private mind that never forgets you. That nobody can switch off, and that knows you better than any system ever has. Because you were the only one who ever trained it. That's not a smaller super intelligence, that's the only one worth wanting." His recommended action: "Take the stranger home and give it a memory." **Stated stance in one sentence:** Intelligence should be owned by individuals through locally held, continuously accumulated personal memory, with any sharing being explicit and consented. # Video Attribution & Provenance Record ## Video Identification - **Title:** *Stop Renting Your AI's Memory — Dylan Couzon, Qdrant* - **Video ID:** `01a15ca7-2254-7a99-a60e-6cec0701a68f` ## Speakers & Roles ### Dylan Couzon (spelled "Cozan" in self-introduction transcript) - **Role:** Speaker / presenter — self-identified as a **developer relations engineer at Qdrant** ("Quadrant" in transcript). - **Context:** Delivers a conference talk titled *"The Frontier Is Coming Home."* He states: *"my name is Dylan Cozan. I am a developer relations engineer at Quadrant."* - **Note:** No other speakers are identified. The closing `>> [music]` cue is non-speech. ## Referenced Third Parties & Sources | Referenced | Claim made in talk | |---|---| | **Karpathy** | Framing quoted by speaker: *"the model is the CPU, the memory the context window is the RAM, and the memory is the disk."* | | **HNSW (2016)** | Cited as evidence memory infrastructure has existed for years. | | **Local embeddings (since 2019)** | Claim that embeddings have run locally since 2019. | | **"Fable and Mythos 5"** | Cited flagship models allegedly pulled by "a government order" last month — third-party claim, not verified in source. | | **YOLO** | Object detection model used in the live demo. | | **Anonymous developers** | Claim some burn *"over $5,000 a month"* of compute on a $200 plan — attributed generally, no named source. | ## Third-Party / Unverified Claims (present in transcript, no source cited) - Sub-$2,500 machines can run "last year's frontier" models. - "Every frontier lab shipped a memory feature" last year. - Smart glasses "already shipping today" were "seen today in this room." ## Promotional Segments — **FLAGGED** - The speaker presents **Qdrant's embedded vector search engine** and its capabilities (local store, quantization, cloud sync, "hive mind" sync). This is a product demo within the talk: - Live offline drone demo using Qdrant (YOLO detections → embeddings → semantic search <1 ms; 92 objects, ~300 vectors, 15 MB footprint). - Qdrant cloud sync offered "out of the box" for opt-in edge-to-cloud sharing. - The talk's conclusion directly encourages adoption of the pattern the sponsor's product implements. ## Technical Assertions Made During Demo (as spoken, no external citation) - Fully offline, no network attachment during demo. - Semantic search (e.g., "coffee table") returned results in under 1 ms. - Cloud sync enables shared memory across fleets ("swarm of drones"). ## Availability Limits - No mention of comment availability, transcript restrictions, or access limits in the source material. ## Evidence Channels Unavailable - Slides and visual demo footage are not included in the transcript; only the spoken narration is available. - No curated metadata beyond the supplied title was provided; speaker identity rests solely on the self-introduction in the transcript. Hi everyone. Thank you for joining me today. I hope everyone is having a great conference so far. Um so my name is Dylan Cozan. I am a developer relations engineer at Quadrant, and the title of my talk today is the frontier is coming home. So what do I mean by that? And somewhere in the next few years, something smarter than all of us is going to exist. We've spent years asking when, but I think when is kind of the boring question. I think the real one is who it belongs to. Because there are two versions of this. In one, super intelligence lives in someone else's data center, and you pay by the token. In the other, it's yours. It learns from your life, your context stays private, and nobody can switch it off, throttle it, or take it away from you. That second version is what I want to talk about today. That's the frontier coming home. A quick thought experiment. What would you build if your assistant actually remembered everything you ever taught it? Every preference, every correction, every dead end for years. That's not a chatbot any anymore. That's a second mind. But the the assistant you're using right now forgets how much all of it the second the conversation ends. Every session starts from zero. We call these things intelligence, and they are. They're just geniuses with no long-term memory. This talk is about fixing that. Not with a bigger prompt, with actual memory. That genius usually exists behind someone's API, in a data center you never see. No, it runs on a on hardware that most of us own. Not the absolute frontier, I'm going to be honest, but shockingly close. A machine under $2,500 today can run what was basically last year's frontier. And open weights models are closing on fast. The gap is shrinking every quarter. You'd expect this to feel like a landmark, uh the frontier on your desk for the price of a laptop. And yet, it doesn't quite land. Because a model on its own is just a brilliant stranger. It's powerful. It's but it it isn't yours. What make What makes it yours is everything it knows about you. And that part hasn't come home yet. So, you can run a frontier class model on your own hardware today. And yet, look at the AI stack that most of us are using today. The compute is rented. The model is rented. The harness is rented. And the parts that's supposed to become you, the thing that should compound over the years, also rented. If we don't fix that, the most powerful technology any technology any any of us will ever touch ends up owned by whoever holds the lease, not the people it was built for. So, let's look at what renting actually costs today. Starts where we starts with where the model leads. Because when you rent it, you don't really control it. This shows up in three different ways. Last month, a government order pulled access to two flagship models, Fable and Mythos 5, for every single customer. Now, compare that to weights you already that are already sitting on your desk. Nobody can reach into your machine and flip those off. Even when the model the model stays up, It doesn't stay put. Providers retire versions and they throttle what's left when the economics gets tight. The model you rely on can quietly get worse or vanish. Weights you own never change unless you change them. People also don't don't write prompts anymore. They're in agents or loops for hours and one task can burn thousands of times more tokens than a chat message. Some developers burn through over $5,000 a month of of computes on a $200 plan. That's a subsidy and those subsidies will ultimately disappear. If you own the computer and the weights, the control comes back to you, but that's still only half the problem. Here's the deeper issue. Owning inference gives you autonomy. Nobody can take the model away. But owning memory gives you continuity. And continuity is exactly what every major lab is trying to package up and sell back to you. Look at last year, every frontier lab shipped a memory feature. Not because the models got smarter, but because they don't carry anything between sessions. That gap is the actual product. A system that starts simple and keeps learning you will always feel smarter than one that starts brilliant but forget everything about you. That company is where intelligence starts to feel personal. So, what actually is memory? It's not a bigger prompt. It's a system with three verbs: write, retrieve, and forget. Same three things your brain actually does. And you know, there's an obvious pushback. Why not just throw everything in a big folder of markdown files? And because retrieving the right memory beats dumping everything and praying the model finds it. Retrieval gives you something a prompt cannot. It is control. You can filter by topic. You can decay by recency and frequency. You can let relevant shift over time the way human memory actually works. And the infrastructure has been ready for years. HNSW in 2016 um embeddings have been running locally since 2019, long before any any model could really use them at scale. The memory side of superintelligence was never the hard part. We were just pointing it at ourselves. So, today the memory that actually knows you lands in one of two places. Either it's stuck in an unindexed file you can't really query, or it's in someone else's cloud. One terms One terms of service change away from being reshaped or moved. Karpathy framed it in a way that's really stuck with me. It's the model is the CPU, the memory the context window is the RAM, and the memory is the disk. RAM is fast, but forgets forgets the moment the session ends. The disk is what remembers you across every conversation. And right right now for almost every AI system on on Earth, the the disk is in someone else's data center. So, you know, today I'm not going to argue which memory architecture wins or is the best one. That's definitely a different talk. But the disk, the parts that holds the actual record of you, should be yours fully and forever. So, this is the piece that I work on, and it's just one way to do it. Um I I work in a vector search engine that embeds directly into your app. It opens look a local store inside your process. You write embeddings with payloads, then query offline in some millisecond time. Same raw score as Quadrant in the cloud, just running where you are. With quantization, a million memories can fit in less than a gigabyte. Small enough to live on a phone or even a Raspberry Pi. Close the app, come back a year later, the memory is intact. Every app gets its own isolated store, and that folder is the thing that actually follows you. So, I'm going to do a quick live demo here. Um so, right now um I have a video recording of a drone like going over a house. So, basically that drone it's it's first time starting. It's it's first time out in the world. It doesn't know anything. And so, this this is actually running live right now, fully offline on my laptop. And I'm using YOLO, which is an object detection model. So, it detects the objects and creates a um label for every items it sees here. Then we take take that picture, we take that label, and we turn that into embeddings to create a live real-time memory for that drone. And And that way we can The drone can remember everything it has ever seen before. And so, here you can see a 2D representation of that vector space with all the memories inside. And memories that are the similar will will be the closest to each other. So, here you can see that we've already recognized 92 different objects, and we have over three 300 vectors. So, that's 300 different representations of those 90 objects. And you can see that the entire memory and Quadrant engine footprint is only 15 megabytes. And then what we can do on this memory? Well, we enable semantic search over those memories. So, you know, if I click here on the coffee table, you can see that in less than 1 millisecond, I was able to pull up every single coffee table that this drone has seen before. And so, you know, we have the image here. We have the first seen times timestamp. We have the last seen timestamp. We have a definition of that coffee table. And and also we can see how many times that table has been seen. So, this is all being run and processed locally. There's no network attachment at all. Um but we also have a capability of cloud sync. So, for example, you know, you if you have like a swarm of drones or robots navigating the world, they can have like a hive mind in the cloud where you know, one system learns something and they can all learn about it. And you know, this is what this was like an example for a drone, but you can apply that to to everything. Just a chatbot, an agent that you call to, a code assistant, and you know, if you create those memories for just a few weeks, the stranger just disappears. It remembers your corrections, your preferences, the dead ends you had to only hit once, and a model that's merely good, but never forgets you and keeps compounding over the years, starts to feel superhuman in practice. Not because the model changed, but because it never stopped learning you. And it's portable. Swap the reasoning model, change the hardware, none of it matters. It all comes with you. If you keep the same embedding model and that folder becomes permanent. A lifetime of context fully yours moving from device to device. But we can go even further than that. You know, this can go way past the laptop or a drone. Point the same memory at your whole day. Where did I leave my badge? What was her daughter's name again? What was on that whiteboard? And what what what was the song that was playing when we met? Every one of those is just a a retrieval query. And you know, you just watched a drone remember objects. Now extend that to everything you see and everything you hear. And you know, that's not science fiction. Those are the smart glasses that are already shipping today and that we have seen today in this room. It's not a smarter you. It's a you that doesn't forget. And right now, none of those memories share uh none of those devices share a memory that you actually own. They could on your own terms because frontier models are already building an index of you today. They have your relationships. They have your habits. They They have your whole inner life. That index is coming whether you opt in or not. The only question is whether you own it or if you're renting access to yourself. And one last idea because owning memory doesn't doesn't mean keeping it to yourself. Edge can sync parts of your store to the quadrant cloud out of the box when you choose to. Picture a family each wearing glasses that record their own day. Separate memories, but they can pull into a hive mind the whole family owns. Mom's day, dad's day, your day searchable together. Now scale that up. Instead of one company's super intelligence serving billions of identical people, you get thousands of small private ones, each shaped by a life, a family, a team. But that same mechanism cuts both ways. A memory can be shared without A memory share can be shared with consent, or it can be extracted without it. That's exactly why it has to start local, and why sharing should always be opt-in, never the default. You should really decide who gets your continuity and your memories. And this is where I'm going to leave you. You know, every everyone here is going to watch super intelligence show up over the next few years. That's That part is basically settled. What's not settled is who it belongs to, and that's the fight worth having right now. Where the archi- where the architecture is still up for grabs. So tonight, don't go home and just try your tool. Trick Take the stranger home and give it a memory. Then picture that memory 5 years from now. Not a chatbot, a private mind that never forgets you. That nobody can switch off, and that knows knows you better than any system ever has. Because you were the only one who ever trained it. That's not a That's not a smaller super intelligence, that's the only one worth wanting. All right. That's all for me. Thank you very much. \>\> \[music\]