Context Portability Context Engineering
Claude Fable 5 put a price on context portability
Key takeaway
AI memory lock-in is the switching cost every major assistant spent 2026 building: the more context a tool holds about you, the more expensive it becomes to leave. Anthropic merged Claude chat and Cowork memory on August 25, 2026, which removes that switching cost between two of its own products, and the reason is instructive. A wall that keeps users in also keeps them out of whatever you build next. The fix Anthropic applied internally, one shared store rather than a copy, is the same fix users need across vendors, and no one is going to ship it for them.
Anthropic spent 2026 making Claude remember you. Memory came to Team and Enterprise, then to Pro and Max, then to the free tier, and each expansion made the same quiet trade: the more Claude knows about how you work, the more it costs you to work anywhere else. That is not a criticism, it is the strategy, and every major assistant is running it.
Then on August 25, Anthropic took a hole out of its own wall. Claude chat and Claude Cowork now share one memory store, reading and writing in both directions. The interesting question is not what users gain. It is why a company would deliberately lower a switching cost it spent a year raising.
The answer is that the wall had turned around and was pointing at Anthropic’s own product.
Every major AI assistant shipped a substantial memory upgrade in 2026, and the reason is retention rather than capability. OpenAI rebuilt ChatGPT’s memory architecture in June, Google introduced Personal Intelligence across its apps, Microsoft brought Copilot memory to general availability in M365, and xAI added persistent memory to Grok. With more than a billion people now using AI assistants regularly, accumulated context has become the most reliable thing keeping any of them in place.
Activant Capital’s research on the memory moat puts the mechanism plainly: the durable asset is “the record of how one business actually works, built up inside a vendor’s product,” and once that record is thick enough, “leaving becomes unthinkable.” Their conclusion is worth holding onto, because the rest of this follows from it: compression drives what people buy in 2026, and ownership drives what they keep in 2030.
This is a better moat than the ones software usually gets. It does not require the product to stay ahead on quality, it costs the vendor nothing to maintain, and it deepens every week the user keeps working. It also has a property nobody talks about, which is that it is not directional.
A switching cost does not distinguish between competitors and colleagues. Context locked inside Claude chat raises the price of moving to ChatGPT, and it raises the price of moving to Claude Cowork by exactly the same mechanism, because in both cases the destination starts cold.
Cowork is the product Anthropic most needs people to pick up. It launched as a research preview on January 12, 2026, aimed at non-technical users doing desktop work rather than code, and the early usage bore that out: of 1.2 million sessions Anthropic reported, more than 90% was non-coding work, with business operations around a third and content creation about a sixth. It is also the product that spilled the coding-agent wars into the rest of the office, which is a large enough prize to reorganize a roadmap around.
Look at how Anthropic has removed friction from trying it. Cowork was folded into plans people already pay for, so nobody has to justify a new line item to their finance team just to run it once. That handles the billing objection. It does nothing about the second objection, which is that a person with eight months of accumulated context in Claude chat opens Cowork and finds a product that has never met them. Bundling makes trying it free. Shared memory makes trying it cheap in the currency that actually matters.
So the August 25 release is an adoption move that happens to be a user benefit. Entries now update continuously during a conversation rather than being summarized at the end, so the two surfaces stay usable at the same time rather than taking turns. Anthropic also exposed the store itself, letting people read, edit, and delete what Claude has retained, with sensitive categories excluded by default. Those are real improvements. They exist because a wall between two of your own products is a tax you pay yourself.
And it worked in a way that proves the general case. When context moves freely between two tools, people use both. Anthropic ran that experiment on its own portfolio and shipped the result. Then it drew the boundary at the edge of the company.
Here is the part that should end the debate about whether export solves this. Anthropic already had an export path. Claude’s memory import has existed since March 1, 2026, built precisely to move what one assistant knows about you into another one. When Anthropic needed Cowork to know what chat knew, they did not use it. They merged the store instead.
That is a company declining its own migration feature the one time correctness mattered to its own revenue. If a copy were sufficient, Cowork ships an import button in an afternoon and every user understands it immediately.
It is worth looking at what these migration tools actually are, because “export” makes them sound more solid than they are. Anthropic’s import works by handing you a prompt to run inside your current assistant, then pasting the output into Claude. The prompt itself asks the source model to output everything in one code block and then instructs it: “Do not summarize, group, or omit any entries.” It closes by asking the model to confirm whether that was the complete set. The transfer format is a request that a system be thorough, followed by asking that same system to grade itself. Google’s Gemini import, launched March 26, works the same way, with a pre-written prompt in the source app and a paste into the destination.
All of it satisfies the letter of data portability. GDPR Article 20 asks that people be able to receive their personal data and transmit it elsewhere, and a text blob technically qualifies. Note where Google shipped its import tools, though: everywhere except the United Kingdom, Switzerland, and the European Economic Area, which is to say everywhere except the jurisdictions where the right to switch is legally strongest.
Export exists everywhere, so nobody can be accused of locking you in. Sync exists nowhere. That gap is the actual lock-in, because the moment you export you have two copies, and from then on they drift. Tell one assistant you changed roles and every other one holds the old answer indefinitely, with nothing marking it stale. As Activant notes, no lab has shipped a cross-vendor export standard, and there is no commercial reason for one to. Most people are already living on the wrong side of that boundary. Among developers using AI, 82% report using ChatGPT, 68% GitHub Copilot, 47% Gemini, and 41% Claude. Those numbers sum well past 200%, which is the whole point: multi-tool is not a transition state people are passing through on the way to picking a winner. It is the steady state, and it is getting more crowded as agents get embedded into products that are not assistants at all.
Every one of those tools holds its own record of you, and none of them reconcile. You decide something with one agent, switch, and the next one does not know, so you re-explain it or it redoes work you already ruled out. The cost is not dramatic on any given day, which is exactly why it persists. It is a tax collected in small enough increments that nobody itemizes it.
Anthropic solved this for the span of its own product line. That solution is not available to a person running Claude, Codex, and two agents built by companies that have never heard of each other, and it never will be, because it requires a shared store and no vendor will host a shared store for its competitors.
The fix Anthropic applied internally is the right fix. One record, both tools reading it, no copy to reconcile. The only thing wrong with it is who drew the boundary.
That is the design Wire starts from. A container holds your context outside any assistant, and every agent you work with reads and writes to that same place through MCP, so a correction written by one agent is what the next agent retrieves without an import step and without either agent knowing the other ran. Adding a fourth tool does not mean a fourth migration, because there was never a copy to migrate. The container is the thing you keep, and the models are what you swap.
This changes which decision is expensive. Under per-vendor memory, the model you commit to today quietly becomes the model you are stuck with, since the accumulated context is the part you cannot move and it grows every week. When the context sits in a container you control, switching models costs you a connection string. That is the difference between choosing a tool and being chosen by one, and it is the reason context portability is worth treating as an architectural decision rather than a feature you look for later.
None of this is an argument against per-vendor memory, which is genuinely useful inside the product that built it. It is an argument about where your record of how you work should live, and the answer that survives the next model release is: somewhere you own. Code has GitHub. Everything else has Wire.
Three questions, in the order they usually matter.
Is this one store or two? A feature that copies is a snapshot with a timestamp on it. A feature that shares is a store both sides read. Anthropic’s August 25 release is the second kind, which is why it is worth paying attention to.
Whose boundary does it stop at? Every memory feature has an edge. Ask where it is, and then ask whether your actual working set of tools fits inside it. For most people it does not, and it fits less well every quarter.
What happens when you add the next tool? This is the question that separates a memory feature from a portable one. If the answer involves running an export, you have found the tax. If the answer is that you point the new agent at what you already have, the context is genuinely yours.
The useful thing about August 25 is that Anthropic answered all three correctly, inside its own walls, and showed everyone what the right answer looks like.
Sources: Claude Cowork finally remembers what you told the app in chat · The memory moat: the last AI switching cost? · The coding agent wars are spilling into the rest of the office · A quote from claude.com/import-memory · Import and export your memory from Claude · Google adds ChatGPT and Claude import tools to Gemini · State of AI agent memory 2026 · AI pair programming statistics 2026 Claude Code, Codex & Cursor
wire-memory writes decisions, corrections, and preferences to a container as you work, keyed to you and your project. Any agent you connect can look them up when it needs to.