ChatGPT Memory v1 arrived in February 2024 as the industry’s first serious fix. Users could say “remember that I’m vegetarian,” and ChatGPT would extract what it treated as a fact, store it in a list, and inject the whole list into every new conversation. The list was viewable and deletable in settings. Khemani calls it a decent first attempt with two flaws.<\/p>\n
First, memory management became the user’s job. “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. The burden of memory management fell to the user.” Second, memories went stale. His own list included the entry “Shlok is going to Bengaluru.” That was true when written, but he now lives in San Francisco, and the old entry is still injected into his chats today.<\/p>\n
ChatGPT Memory v2 in April 2025 replaced the fact list with a running profile. Every few days, the system reviews all of a user’s conversations, extracts what it thinks matters, and updates the profile in the background \u2014 a process researchers call dreaming. The updated profile then loads into every new conversation. Think of a token as a chunk of a word. Khemani’s profile runs about 4,000 tokens across 16 sections, roughly the length of a long article. The memories are dense, keyword-like clues rather than full sentences, on the assumption that a frontier model can expand those clues mid-conversation. V2 moved the burden off the user, but staleness survived. ChatGPT’s profile fabricated the Turkey trip from chats in which he merely weighed Thailand versus Turkey.<\/p>\n
Claude came at the problem from the opposite direction. In August 2025, Claude’s v1 memory had no profile at all. The model received two tools: search past conversations by keyword or topic, and search by time period \u2014 for example, “what did we discuss at the start of November of 2025?” Every conversation started fresh, and retrieval happened only when needed. Khemani published a post titled “Claude’s memory architecture is the opposite of ChatGPT’s” on September 11, 2025. It hit the front page of Hacker News \u2014 and that same day, Anthropic shipped Claude’s v2, which added a running profile. The implementation was deliberately different: a raw profile visible in settings, about 1,000 tokens rather than 4,000, complete sentences instead of dense keywords, updates every 24 hours instead of every few days, and explicit user edits that trigger the profile to be rewritten. Claude’s memory has not changed since September 2025.<\/p>\n
ChatGPT answered twice in 2026. First, it added a tool to retrieve summarized context from past conversations on the model’s own query, effectively adopting Claude’s on-demand retrieval. Then, in June 2026, it made the profile visible in a partial form: users see an LLM-generated summary of the profile rather than the raw profile itself. Users can now request edits, and OpenAI deprecated the old v1 fact list.<\/p>\n
| Attribute | ChatGPT | Claude |
\n|—|—|—|
\n| Profile size | ~4,000 tokens across 16 sections | ~1,000 tokens |
\n| Memory style | Dense, keyword-like clues | Complete sentences |
\n| Update cadence | Every few days | Every 24 hours |
\n| User visibility | Hidden until June 2026; now an LLM-generated summary | Raw profile visible since September 2025 |
\n| User editing | Edit requests supported since June 2026 | Edits trigger re-synthesis; edit history managed in UI |
\n| Past-conversation retrieval | Tool added in 2026 | Keyword\/topic and time-period tools since August 2025 |<\/p>\n
By mid-2026, the two products are much closer than they look. Both run a profile plus retrieval, and both are now nominally visible and editable. But the details are still products of different judgments.<\/p>\n
<\/p>\n
### There is no single way to do memory<\/p>\n
For a while, many people in the field \u2014 Khemani included \u2014 assumed RAG would be the obvious route: break conversations into chunks, turn those chunks into numerical fingerprints, store them in a searchable database, and run a similarity search whenever the model needed context. Khemani says neither ChatGPT nor Claude does anything close to that. The two products evolved independently, and although the broad architecture converged, the implementations remain different.<\/p>\n
A broader scan makes the point even stronger.<\/p>\n
| System | Memory mechanism |
\n|—|—|
\n| ChatGPT | Running profile + conversation retrieval tool |
\n| Claude | Running profile + keyword\/topic and time-period retrieval |
\n| Gemini | Running profile with per-memory timing logs |
\n| Agent ecosystems (Claude Code, OpenClaw, Hermes) | Markdown files, heartbeat processes, knowledge bases, skills |<\/p>\n
“The biggest lesson for me is that there is no single way to do memory,” Khemani said. That has a direct strategic consequence: memory cannot be treated as a bolt-on component. “Memory cannot be outsourced. If you’re a serious team, you do not outsource memory. It is something that you build alongside your product.”<\/p>\n
The evidence is visible in the market. The top consumer products across categories all have some form of memory, and none of them outsource it. They all build it in-house, not because there is a standard blueprint, but because the system has to evolve with the product.<\/p>\n
<\/p>\n
### Memory is a function of compute<\/p>\n
A running profile carries two separate costs. The maintenance cost comes from how often the system updates the profile and how much compute it spends on each update. The serving cost comes from putting that profile into the context window of every single conversation \u2014 a longer profile makes every message more expensive. The two costs trade off directly.<\/p>\n
| Design knob | Unconstrained ideal | ChatGPT | Claude |
\n|—|—|—|—|
\n| Update cadence | Hourly, or after every conversation | Every few days | Every 24 hours |
\n| Profile size | 400,000 tokens | ~4,000 tokens | ~1,000 tokens |
\n| Update compute | Fleet of Opus sub-agents | Lower per update | Higher per update |
\n| Dominant cost | \u2014 | Serving | Maintenance |<\/p>\n
“Memory is a function of compute. You have to really think about how much compute you want to put into memory,” Khemani said. An unconstrained design would update the profile every hour, or after every conversation. It would use a fleet of Opus-level sub-agents to think about each update. It would keep 400,000 tokens of context on every user. The reason nobody ships that is not that it would be useless. It is that the industry is GPU-constrained.<\/p>\n
Those constraints are not abstract. Separate from the user-memory design Khemani studies, the physical memory chips that power these systems are in a historic shortage. As of mid-2026, DRAM prices have climbed back toward levels last seen roughly two decades ago, and Micron executives have warned that calendar 2027 will be even tighter than 2026. The compute dials Khemani describes are not just product preferences; they are responses to a genuine physical bottleneck.<\/p>\n
As compute gets cheaper over time, expect ChatGPT-style profiles to grow and Claude-style update cadences to accelerate. That makes profile length and update frequency a useful leading indicator of where the industry thinks its cost structure is heading.<\/p>\n
### Continual learning is already here \u2014 outside the weights<\/p>\n
One of the bigger reframes from Khemani’s analysis is that continual learning is not a future research goal. It is already running in production, just outside the model’s weights.<\/p>\n
The running profile starts as what the model believes about the user. It is added to every conversation. Conversations generate new information. The dreaming process turns that information into an updated profile. The updated profile shapes the next round of conversations. That loop repeats indefinitely. The only twist is that the model’s weights do not change; the learning happens in a separate text layer.<\/p>\n
“Continual learning is already here,” Khemani said.<\/p>\n
The harder question is whether that loop moves into the weights themselves. A model that truly learned from every user would be far more personalized. But training or fine-tuning weights for one person is expensive. Enterprise continual learning can amortize the cost across employees and customers. An individual’s model has nobody to share the bill with. Khemani says he cannot yet answer three things: whether each person gets their own self-learning model, what data bootstraps the process, and who pays. He points to Guan’s essay “Guardian Angels” as a reference for the economics of a one-model-per-person future.<\/p>\n
### The context ceiling: where memory stops and frustration begins<\/p>\n
Even the best memory system, with infinite compute and individual-level learning, is capped by one thing it cannot control: how much context it is allowed to gather about the user.<\/p>\n
The Thailand\/Turkey example is the exhibit. The source conversations contained only deliberation between two candidate destinations. The actual decision happened in an in-person conversation. Evidence of the outcome existed in email \u2014 Thailand flight and hotel bookings. But even if ChatGPT is connected to his email account, it does not reason over email or update his profile from it. So the contradiction went unresolved, and the false entry kept being served. The detail that bothers Khemani most is not the staleness itself. It is that the system does not notice the conflict and is not curious about the gap.<\/p>\n
“It’s not a technology problem. It’s a product problem,” he said. No fundamental limitation of large language models prevents progress here. The products are simply not designed to gather and reconcile context.<\/p>\n
His personal stack makes the point concrete. He runs chatbots, assistants, vertical-specific applications, agents, and hardware devices. Every one of them builds its own separate memory of him. None of those memories are shared. Every life change has to be communicated individually to each product. Meanwhile, his richest context sources \u2014 email, calendar, photos \u2014 are not reasoned over by any of them.<\/p>\n
“None of this feels like 2026. … When will personal AI feel like personal AI?”<\/p>\n
### What to watch<\/p>\n
After three years, the pattern looks like a commodity: profile plus retrieval, built in-house, tuned to a compute budget. That shifts competitive differentiation to the two things major products still do poorly \u2014 context gathering and the economics of individual-level learning.<\/p>\n
| Development | Why it matters |
\n|—|—|
\n| A major assistant begins reasoning over email, calendar, and photos to build and correct its profile | That is the exact gap behind the Thailand\/Turkey failure \u2014 the difference between a profile fed by chat and a profile that reflects a life |
\n| Per-user weight-level continual learning finds an economic model | Enterprise amortization already works; the individual case does not. Guan’s “Guardian Angels” is the reference point |
\n| Memory becomes a portable, user-owned layer | If every product re-learns a user from scratch, a shared memory substrate would change competitive dynamics |
\n| The compute dials move | As serving costs fall, profiles should grow and update cadences should accelerate |
\n| Visibility and control expand as a trust surface | Claude led with a raw profile in September 2025; ChatGPT followed with a partial LLM-generated summary in June 2026 |<\/p>\n
That is why Khemani treats memory not as a feature but as the substrate of the entire human-AI relationship. He expects to talk to AI systems for the rest of his life, which makes memory “something that’s going to be important for the rest of human history.” For investors, the signal is that the easy differentiator \u2014 having a memory at all \u2014 is already table stakes. The next winners will be whoever can pull in context from outside the chat window and find a viable economic model for making the model actually learn from each user. There is also a more immediate signal: the compute dials Khemani tracks are already being pushed by a historic shortage in physical memory chips, so profile sizes and update frequencies will reveal how strained the underlying hardware really is. Three years in, he says, “there’s so much left to build.”
\n ]]><\/p>\n","protected":false},"excerpt":{"rendered":"ChatGPT Memory v1 arrived in February 2024 as the industry’s first serious fix. Users could say “remember that…\n","protected":false},"author":2,"featured_media":138131,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[53,580,182,2798,2408,132,68215,43109,585,157,576,68214],"class_list":["post-138130","post","type-post","status-publish","format-standard","has-post-thumbnail","category-openai","tag-anthropic","tag-chatgpt","tag-claude","tag-claude-code","tag-gemini","tag-google","tag-guardian-angels","tag-hermes","tag-micron-technology","tag-openai","tag-openclaw","tag-shlok-khemani"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/138130","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=138130"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/138130\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/138131"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=138130"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=138130"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=138130"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}