A personal agent’s most valuable output may never appear in a document. Its real asset is the accumulated understanding of how its user works, including preferences, recurring exceptions and the significance of particular relationships. If that understanding exists only inside one product, every useful interaction can deepen dependence on its supplier. The purchase decision therefore includes an undeclared question: can the agent’s understanding leave with the customer?
The AI Daily Brief episode behind this discussion reports that getting value from a personal agent requires deep context, setup, tool access and account access, and says later switching could be costly. The concern deserves attention, but app connections are only the visible layer of the problem.
The Model Context Protocol is an open protocol for connecting language-model applications with external data sources and tools. Its specification covers resources, prompts and tools that servers can offer to clients. That can reduce the work required to connect another agent to the same systems. Memory sits elsewhere because access to a customer relationship system does not give a new agent the judgement accumulated by its predecessor.
Anthropic’s help page says individual Claude users can export account information and chat history. The same page says exported data cannot be imported into another personal Claude account. This example is limited to one product policy, so it cannot establish the condition of the wider market. It does show why the existence of an export button says little about practical mobility.
An export is an exit artefact. Usable portability requires a receiving system that can interpret the material, preserve meaningful distinctions and identify what should no longer be trusted. If a file contains conversations without the preferences inferred from them, the customer may possess the record while losing the benefit. Without a workable import path, data can leave while operational context remains stranded.
This changes what procurement should examine. An integration checklist answers whether the agent can reach email, files and business applications today. A memory-portability assessment asks whether a replacement could reconstruct useful context without replaying months of work. The relevant unit is the cost of restoring competent assistance, including review and correction.
Consider a hypothetical account director whose agent has learned which customers expect detailed follow-up, which draft language needs legal review and which internal requests can wait. If the incumbent exports only chat transcripts, the director or employer must determine which exchanges contain durable instructions and which reflect obsolete circumstances. If a competing agent cannot import that knowledge safely, changing supplier creates retraining work and a period of weaker assistance. The nominally better product could then be the more expensive commercial choice.
Before approving a substantial deployment, give each shortlisted supplier a synthetic user history containing preferences, corrections, expired instructions and conflicting examples. Require it to export that history and explain how another system could distinguish durable memory from raw conversation. Then test the proposed transfer into a separate environment rather than accepting a file as proof. Success means the receiving agent can use relevant context while exposing uncertainty for human review.
Commercial terms should address the same issue. Buyers should specify the timing, format and completeness of memory exports, along with access after termination. They should also separate business records from inferred preferences because the receiving system may need different controls for each. The distinction matters when an agent’s conclusion is useful but the basis for that conclusion is unclear.
Portability carries its own risk. An export containing sensitive relationships, behavioural patterns and delegated authority could be more revealing than a folder of ordinary documents. Controlled portability therefore needs authentication, selective transfer, deletion controls and an audit trail suited to the sensitivity of the memory involved. The objective is a credible exit path without creating a convenient package for misuse.
Businesses can also reduce dependence through architecture. Where practical, authoritative instructions and important decisions should remain in systems the business controls, with the agent retrieving them as needed. If the agent keeps private memory, treat it as a replaceable working layer rather than the sole corporate record. This approach can make replacement less destructive even when vendor-to-vendor transfer remains imperfect.
The vendor with the richest memory can become difficult to displace without imposing any contractual lock-in. If personal agents gain deeper access and deliver more useful work, that quiet dependence will become more commercially important. A market can support standardised tool connections while leaving accumulated judgement trapped inside proprietary products. Until buyers demand importable memory, export rights will offer the appearance of control. Ask a prospective supplier to demonstrate the first useful hour after migration. The answer will reveal more about switching costs than another page of features.
sources
Prepared with assistance from a Useful Engine agent workflow. Factual and editorial checks are part of the publishing process.