The rented mind: when intelligence is cheap, memory is the lock-in

Overview

This weekly edition is about a single distinction that quietly decides who holds the power in the AI economy — and why, once you can see it, you read every product announcement and every internal rollout differently.

Why now: within eight days this month, both frontier labs took the step that reframes everything. Anthropic confidentially filed for an IPO on 1 June 2026, reportedly near a US$965 billion valuation; OpenAI followed about a week later, with a filing reported to target a valuation around a trillion dollars. The valuations aren’t the interesting part. What public investors are being asked to believe is: that these companies can make raw intelligence cheap enough to serve at vast scale, and own the layer of work that sits on top of it — fast enough that companies rent the whole system instead of building their own.

That second half is the whole game. By the end you’ll have the distinction that organises the bet — model versus harness — the reason one part of a harness matters more than all the others, and a one-sentence test you can apply to any AI system you depend on, yours or your employer’s.

The content

Start with the distinction, because everything follows from it. A model is raw intelligence — the thing you buy by the token. A harness is everything that turns that intelligence into work: the files it can see, the tools it can call, the permissions it holds, the memory it keeps, the checks on its output, the routing between a cheap model and an expensive one. Birgitta Böckeler, writing on Martin Fowler’s site in April 2026, frames it through the shorthand the field has settled on — Agent = Model + Harness (a formula she credits to LangChain) — and defines the harness as “everything in an AI agent except the model itself.” A model gives you intelligence. A harness gives you work.

Now add the economics, because they decide where the value goes. The cost of a given level of AI is collapsing. Sam Altman, in early 2025, described it as falling “about 10x every 12 months”, and noted that Moore’s law “changed the world at 2x every 18 months; this is unbelievably stronger.” Epoch AI’s measurements are more conservative but point the same way — a several-fold cost reduction per year for a fixed level of capability, with the caveat that the rate may slow. When an input gets cheap that fast, it stops being where the value lives. Raw intelligence is being commoditised in real time, and the labs know it. Cheap is the plan, not the accident.

So value migrates to the harness — and both labs are betting their valuations on owning it. That is what the new “deployment” arms are for: in May 2026 OpenAI stood up a multi-billion-dollar deployment company and bought a forward-deployed-engineering firm to embed its people inside customers — days after Anthropic had launched a parallel venture of its own. The pitch is not “rent our API.” It is “let us rebuild your workflows around our system.” Read alongside the IPO filings, the strategy is explicit: be the operating layer, not the supplier.

Here is the turn the headlines miss. A harness has many layers, and most of them are replaceable — you can swap a router, rewrite an eval, repoint a permission. But one layer is different, because it accretes: memory. Every day a system runs, it learns more about how you work — and the longer it runs, the more it would cost you to start again somewhere else. New America’s Open Technology Institute named the mechanism in late 2025: a “context flywheel”, where “the more context an agent collects and the more personalized it becomes, the harder it is to leave” — and pointed to the warning that “context, not model performance, is the true source of monopoly power.” The model is the part everyone competes on. The memory is the part that quietly takes you hostage.

Which gives you one question that cuts through all of it — call it the exit test: if you left tomorrow, could you walk out with what the system has learned about you, in a form you could use somewhere else? Today, mostly, you cannot. Every major assistant lets you “export your data” — but what comes out is a dump of your past conversations, not the usable memory the system has distilled from them. The two are not the same. As of June 2026, OpenAI’s newest memory architecture synthesises what it knows about you automatically, in the background — and OpenAI’s own help documentation concedes the memory summary it shows you “will not include everything that ChatGPT remembers.” Google’s Gemini added a feature this March to import your memory from competitors — but the path back out is still a chat-log export, not a portable memory file. Anthropic’s Claude is, to its credit, the one assistant offering an explicit two-way import and export — and even there, what you get is plain text you prompt it to write out, flagged “experimental” on the way back in. Across the board, the memory layer is something you can sometimes read; rarely something you can leave with.

That gap matters most exactly where the stakes are highest: inside organisations. Picture the common enterprise pattern — a capable frontier model wrapped in an internal layer, pointed at the company’s own documents through a retrieval connector, quietly accumulating a memory of how the business actually works. Now put it to the exit test. Who owns the index those documents built? Could the organisation extract the accrued memory and run it against a different model next year? For every major provider, the first answer is undocumented and the second is effectively no: the source files stay yours, but the derived memory — the part that compounds, the part that is actually valuable — lives in the vendor’s store. An organisation can pour years of institutional knowledge into a harness it does not own, and find that the asset it built cannot leave with it. For knowledge management, that is not a feature gap. It is the whole question: memory is the one input you cannot re-derive after you’ve gone, and the one most likely to be held where you can’t reach it.

None of this is hypothetical, and it is already drawing a response. A layer of provider-independent memory tools has grown up specifically to keep the accruing context outside any single lab: open frameworks such as mem0 and Letta that run against any model; local-first memory you own and can read from any client, like OpenMemory over the Model Context Protocol; and, as of June 2026, a W3C community group whose stated mission is a specification for agent memory “portable across vendors, models, agent frameworks, and tool ecosystems.” These exist because the problem is real. Their shared idea is simple: if the memory layer is yours, the labs compete to serve it — they are suppliers. If the memory layer is theirs, they are the operating layer, and you are renting your own institutional knowledge back from them.

Two honest caveats, because the strong version of this argument over-reaches. First, portability is not free interchangeability: the same model can behave differently across providers, abstraction layers can limit you to the lowest common denominator, and most organisations should buy far more of their harness than they build. “Own your memory” is not “build everything.” Second, vendor-native memory is genuinely useful, and refusing it on principle is its own kind of waste. The narrower claim is the one that survives scrutiny: whatever else you rent, keep a copy of the memory layer you control — because it is the one part of the harness that compounds, and the one you cannot reconstruct once you’ve walked out the door.

When intelligence was scarce, owning the smartest model was the position. When intelligence is cheap, the model is something you rent by the token — and the only thing in the system that isn’t fungible is what it has learned about you. Whoever holds that holds the exit. So the question worth asking of every AI system you lean on, long before you ever want to leave one, isn’t “how good is the model.” It’s “whose memory is this?”

Additional reading

• Harness Engineering — Birgitta Böckeler, martinfowler.com, April 2026 — the cleanest statement of “Agent = Model + Harness” and why the hard engineering now lives in the harness.
• AI Agents and Memory — New America / Open Technology Institute, Nov 2025 — the “context flywheel” and the argument that context, not model quality, is the real source of lock-in.
• Three Observations — Sam Altman, Feb 2025 — “the cost to use a given level of AI falls about 10x every 12 months”; pair with Epoch AI’s more conservative measurement (Feb 2026).
• Anthropic files for IPO — Fortune, June 2026 and OpenAI files for IPO — CNBC, June 2026 — the public-market bet that frames the whole piece.
• OpenAI launches the Deployment Company — Constellation Research, May 2026 — the labs racing to own the work layer, read by analysts “through the lens of lock-in.”
• Memory FAQ — OpenAI Help Center, Import your memory into Gemini — Google, March 2026, and Import and export your memory from Claude — Anthropic — the live state of memory portability across the three major assistants (a chat-log “data export” is not the same as a portable memory layer).
• AI Agent Memory Interoperability Community Group — W3C, June 2026 and Introducing OpenMemory MCP — mem0 — the provider-independent memory layer taking shape in response.

Editor’s note

The reason this distinction is worth dwelling on is that almost everything consequential follows from it. Raw intelligence is already close to a commodity — for most work a user can move between providers without much cost — so the question that will shape the next phase is not whose model is best, but who comes to own the harness: the model providers, their customers, or the intermediaries now forming in between. That outcome is far from settled, and working life will look materially different under each version of it. For knowledge workers, one under-appreciated part of the harness matters more than the rest: memory and context management. It is where the value accumulates, it is the hardest thing to recover once another party holds it, and it is being decided now, largely out of sight. It deserves your attention.