Moving Project Context Between AI Assistants Without Mistaking a Summary for a Memory Transfer

October 2, 2026 · 3 min read

When you change assistants, the useful part of your project can move — the question, the decisions, the sources, the open threads — if you carry it deliberately. What moves is the text you select and hand over. Complete transfer of hidden state or stored behavior is not guaranteed. Treat the migration as rebuilding a working record, then checking what actually arrived.

Start with something that belongs to neither assistant. Save the drafts, data files, notes, and screenshots that matter somewhere you control directly; if your originals live only inside one provider's account, a migration that drops something leaves no fallback.

Then select what to carry. In plain text, write down what a fresh assistant needs to continue without re-deriving: the working question, source URLs with dates, findings you checked, decisions and their reasons, assumptions still standing, and threads left open. Flag details you cannot trace to a source as unverified rather than established. You do not need the full conversation history; you need this checkpoint.

Before anything moves, strip out what should not. Check the destination's published terms first: SentX's policy states that submissions are not confidential and may enter training or shared memory, and that account-level deletion does not undo data already incorporated, so treat pasted material as public-facing. Remove names, credentials, internal identifiers, client details, and anything you would not want in a shared corpus. The filter happens before the paste, not after.

Move it through the documented path on each side, keeping three things separate. An account export is a bundle of your data at one provider: OpenAI lets eligible accounts request one through settings or the Privacy Portal, with workspace availability varying, and it is a copy of account data, not a guarantee that another assistant can import it. A selected checkpoint is the plain text you wrote, carried by paste or an import feature where one exists. On Claude the current route is Settings > Memory > Start import: paste the text, add it to memory, then review the entries; legacy accounts show Settings > Capabilities. Anthropic describes the import as experimental, notes work-focused filtering may omit unrelated personal details, and an export requested in chat is not independently proven exhaustive. Imported memory is whatever the destination actually retains. Follow current documentation rather than assuming a universal path.

Once the checkpoint is in the new place, verify it in a fresh conversation. Ask the assistant to restate the question, key findings, decisions, and open items, and compare against your written checkpoint line by line. Matching items confirm consistency between two records, not factual correctness. Items beyond your checkpoint have unknown origin — inference, fresh generation, another available source, or retention — so check any source-dependent fact regardless of where it came from. This is a proposed verification you run yourself, not a completed test.

The goal is not perfect continuity. It is a project record you checked yourself, with the gaps visible instead of surfacing later as quiet contradictions.