What AI Memory Portability Means for Enterprise Control

Our analysis is that portable AI memory should be treated as a governed context transfer, not a complete migration of an enterprise assistant. The buying question is what moves, who authorises it and which controls survive. Exporting useful preferences can be valuable without reproducing an organisation’s conversation archive, project boundaries or retention arrangements.

A memory export is not a workspace migration

Anthropic describes memory imports as experimental and warns that imported memories may not always be successfully incorporated (Anthropic’s import guidance). Our recommendation is to inspect the result with a small, approved set of preferences before moving sensitive business context.

Anthropic’s earlier memory introduction described professional context, team processes and project details as intended uses (The original product explanation). That describes the value proposition, not proof that another provider will preserve every instruction. Buyers should test whether important constraints are still followed after transfer.

Project boundaries deserve their own acceptance check

Claude’s current guidance says project memory can be referenced outside the project when memory is enabled, and identifies an account-memory control for keeping a project separate (Current Claude documentation). This makes the boundary a configuration question rather than something we would infer from the word project.

OpenAI explains that saved memories are stored separately from chat history and that deleting the original chat does not automatically delete a separate saved memory (OpenAI’s memory documentation). We recommend separate tests for removing a conversation and removing the information retained from it. A visible chat deletion should not stand in for both checks.

Export authority belongs in the buying specification

Claude Team and Enterprise members can export their own data once an organisation Owner or Primary Owner enables that option (Claude’s individual export instructions). Before migration, a buyer should agree who can authorise a transfer and which records are allowed to leave.

Organisation-wide Claude exports are available only to Team and Enterprise Primary Owners (The organisation export guidance). OpenAI says self-service export is not available for ChatGPT Business, Enterprise or Healthcare workspaces (ChatGPT’s export instructions). Our assessment is to verify the actual exit procedure for the purchased plan rather than borrow assumptions from consumer accounts.

Private modes and retention are different controls

OpenAI says a temporary chat can use existing memories with personalisation enabled, but neither personalisation option creates or updates memories while the chat remains temporary (Temporary Chat documentation). We would test reading existing context separately from writing new context.

Anthropic says incognito conversations follow retention rules, including an Enterprise organisation’s custom retention setting (Claude’s incognito guidance). For procurement, our interpretation is that a private-session label should trigger questions about storage and access, not assumptions that no record exists.

Measure the handover, not the export button

OpenAI’s data-controls guidance says available controls depend on the account’s plan and workspace settings (The data-controls documentation). We recommend a handover checklist covering authorised content, excluded material, memory deletion and destination access. Those are proposed acceptance criteria, not claims that providers share a common migration standard.

A useful pilot would compare behaviour before and after an approved transfer, document missing preferences and confirm who can reverse the import. The commercial objective should be a demonstrable exit process with known limitations. We would not equate that with lower switching costs until the organisation has measured the work involved.

Related reading: self-hosted AI economics, open-source platform choices, enterprise trust, privacy evaluation, measuring AI returns. These links provide wider business context, not independent evidence for the sources examined here.