Lab
Why a handover needs more than one search
Short answer
A handover connects information that may use different words. One search identifies an asset; another finds the decision explaining its unusual treatment. The Nawame Lab studied this form of multi-step retrieval, including the danger of losing essential evidence along the way.

The missing piece may be an intermediate fact
Consider a fictional delivery that requires an additional inspection. One note names the affected batch. Another, filed under a supplier, explains the incident behind the rule. The supplier’s name may never appear in the initial question.
Adding synonyms cannot solve this dependency by itself. The assistant needs to read the first result to discover what to search for next. An initial campaign investigated this problem. An initial attempt generated queries that were too long. Correcting them helped, but the experiment did not meet every target for questions requiring two retrieval steps.
Let the question determine the next step
A second campaign explored adaptive retrieval: continue when evidence requires it, then assemble the results. Its benchmark used a fictional organisation with teams, restricted folders and changing responsibilities. Several variants were compared on unseen questions using two response-model families.
The experiments found improvements in connecting sources and identifying the current owner of a case. Handovers remained fragile. An overly selective merge could retain the strongest individual results while dropping evidence needed for a complete list.
A third campaign therefore separated two tasks: answering a precise question and preserving the breadth of a case. A handover needs evidence of commitments, subsequent changes and closure. A single relevance ranking does not automatically preserve those relationships.
Sometimes the missing context is metadata
These studies also exposed the importance of authorship. A note says, “I promised them.” Retrieval may find that sentence, yet the response model cannot reliably identify who made the commitment if the author is absent. Showing provenance and searching someone’s writing address a problem that semantic similarity alone cannot solve.
History matters for the same reason. A superseded decision may be unhelpful when asking what to do now and essential when asking why a policy changed. Relevant retrieval depends on the question’s time frame as well as its vocabulary.
Harder tests narrowed the conclusion: scope and limits
The next campaign introduced more difficult cases: commitments spread across records, closure recorded elsewhere and deeper revision histories. Gains were smaller than in the previous one. Temporal measurements in these campaigns later received an erratum concerning version dates; the comparison must be read with that correction. A benchmark score cannot become a general product-quality promise merely because an earlier experiment looked encouraging.
The organisations were synthetic. These scenarios establish neither guaranteed completeness for real cases nor an advantage independent of the receiving model. They help isolate mechanisms worth testing further. An assistant’s ability to use evidence also remains distinct from the retriever’s ability to find it.
What this means for memory design
A useful handover separates established facts, open issues and conclusions that available evidence cannot support. It exposes gaps instead of filling them with a plausible narrative. Every retrieval step must remain governed by the reader’s access rights.
For Nawame, the consequence is to treat retrieval as an evidence-led process. More search rounds do not mean better results. Fidelity, coverage and cost must be assessed together. Reading memory at a specified date and preserving reasons for decisions address other parts of the same problem.
Related reading: How to move project context from one AI to another and Switch AI Assistants Without Losing the Why Behind Decisions.


