Lab
An optional parameter can hide the right memory
Short answer
When an assistant says that no memory covers a subject, the problem may lie in its tool call. The Nawame Lab encountered a case in which automatically filled optional parameters emptied retrieval. The information existed; the query prevented the model from seeing it.

A low score appeared to indict the model
During a comparative campaign, one model performed poorly while others accessed the same memories normally. Looking only at the answers could have suggested a weakness in reasoning or search.
Tool traces revealed a different cause. The client populated optional fields, including an author filter, with values unrelated to the question’s intent. This could cause an error or narrow retrieval without a meaningful reason.
A fictional example makes the effect clear: searching for an asset while filtering for a person who never wrote about it excludes the relevant notes. Reporting “no information available” then turns a malformed request into an apparently factual conclusion.
The contract is part of memory quality
An optional parameter can be absent, empty, invalid or deliberately requested. These states differ. A tool needs to explain what it applied and why a search returned nothing.
Nawame’s MCP audit studied these distinctions through real calls from several models and a deterministic test matrix. It led to an explicit treatment of empty values, visible reporting of certain ignored relevance parameters and signals for limited result sets.
The difference between a relevance filter and an authorisation scope is crucial. Adjusting a search filter must never expand the caller’s rights. An invalid access scope is not permission to search somewhere else. Robustness must operate inside the existing access boundary.
There is more than one kind of empty result
A query may find nothing because no accessible source answers it. It may instead use a filter with no match, target an unresolved folder or return only part of the available results. Distinguishing these cases lets the assistant decide whether to reformulate, continue or explain a limitation.
Pagination is particularly important for handovers. A short list can appear complete if the tool gives no indication that more results exist. Answer completeness then depends on an interface signal even when retrieval itself works correctly.
Test the whole path
The discovered defect led to invalidating and rerunning the affected executions, with the correction documented. Comparing models without inspecting their calls would have attributed a tool-chain effect to reasoning quality.
This does not make every model interchangeable. It establishes the conditions for a fair comparison: the same sources, consistent access rights, correctly interpreted parameters and visible limits. A more emphatic field description cannot substitute for a robust contract.
The consequence for integrations: scope and limits
For Nawame, useful memory includes an interface that authorised assistants can read correctly. The contract must accommodate client differences without inventing intent or hiding errors. Writes require additional care: omitting or interpreting a field can change a decision, so reading rules cannot automatically be transferred to writing operations.
This research complements portability and access rights and experiments on evidence that models fail to use. Before asking for a stronger model, check that the path to memory faithfully expresses the request.
Related reading: How to test non-writing and A good memory must know when to retrieve nothing.


