In detail
Start with the records you actually need
A move between assistants does not require copying every conversation. Choose the preferences, task state, facts, and procedures the next application needs. The memory-types explainer helps separate those uses. Keep the record’s source, owner, date, access scope, and superseded status alongside its content.
Treat import and continued use as separate tests
In episode 77 at 4:41, Kimberly Logan explains why an import alone does not solve workflows that continue using several models. Test the initial transfer, then change a fact in the source and test the next retrieval. Decide which system owns the record and how conflicts are resolved.
Kapil Chhabra makes the institutional version of the argument in episode 75 at 1:19:55: a customer should be able to take its context away from its vendor. He describes export and repository-backed context in WisdomAI. An export path is useful evidence of an exit route; inspect what another application can actually do with the exported content.
Write down the transfer contract
| Question | A small acceptance test |
|---|---|
| Can another application read the records? | Import a representative sample and inspect the resulting fields, including provenance. |
| Does it retrieve the right version? | Update a fact, then ask a task that depends on the new value. |
| Are access boundaries preserved? | Include an unauthorized tenant’s record and verify that retrieval excludes it. |
| Can a preference be withdrawn? | Withdraw one at the source and check each consumer after the update. |
| Is approval being mistaken for memory? | Resume a task with saved state and verify that a new consequential action still needs its own authorization. |
Recheck behavior after switching models
Use the same small task set before and after the move. Inspect what was retrieved and how the agent used it. Compare task outcomes, stale-record use, access failures, and latency. Choose acceptance criteria for your application before calling the transfer successful.
An interface that exposes stored records does not establish that different consumers interpret them alike. Ask separately how updates travel, how access is enforced, and what happens when the source is unavailable.
Keep the sourcing visible
Walrus sponsored the Kimberly Logan interview; WisdomAI sponsored the Kapil Chhabra interview. Their descriptions explain their product approaches. The checklist here is an editorial proposal for testing portability, not independent verification of either product. Logan also distinguishes integrity from truth at 23:45: an unchanged record can still be inaccurate.
For the underlying distinction between active context and persisted records, read Prompt vs. Context vs. Memory Engineering.
Hear it from the guest
“The problem is not solved with an import function either.”
“Context is the IP of the company and it needs to stay with the company.”
Quotes lightly edited to remove filler words.
From the conversation
This explainer is drawn from these episodes — each carries its full transcript.
Concepts in this explainer
Agent Memory PortabilityAgent MemoryLatencyPrecision and Recall