Lab
Reliable memory keeps sources, dates and disagreements
Short answer
Finding a sentence is not the same as knowing whether it still applies. Reliable memory needs to preserve who said what, where it came from, when it became relevant and what authority the statement carries.

An accurate quotation can still mislead
Consider a fictional delivery discussion. During a meeting, a manager says a shipment should arrive on Friday. The minutes record that estimate. An assistant later retrieves the sentence correctly and tells a colleague that Friday delivery is confirmed.
Retrieval succeeded, but the information changed status. An estimate became a commitment. Adding a citation does not undo that shift.
The Nawame Lab separates several questions. Is the source authentic? Does it support the claim being made? Is the reported statement true in the outside world? These questions interact, but none settles all the others. A faithfully quoted rumour is still a rumour. A decision attributed to the right person may have been superseded.
Dates answer different questions
A note’s creation date records when information entered the system. It does not necessarily tell us when a decision takes effect. A procedure may be announced now and apply next month. A correction written later may concern an earlier event.
In the delivery example, a subsequent confirmation could replace the original estimate. But the newest document is not automatically the most authoritative. It may contain a question, a copied message or an informal comment from someone who cannot approve the change.
Useful retrieval therefore needs to preserve relationships between statements. “This estimate was replaced by a confirmation” conveys something different from “this note was written later”. When something became known and when it applies should remain distinguishable.
A disagreement can be the information that matters
When people disagree, compressing their statements into a single confident sentence can hide the work still required. If purchasing expects delivery on Friday while operations needs the equipment on Thursday, both statements need to remain visible with their sources. The useful response may be a question to resolve, rather than an invented consensus.
The Lab programme treats contradictions, revised values, corrections and attribution as evaluation scenarios in their own right. It asks whether a representation preserves the distinctions needed to continue the work, even when the original records are condensed.
This leads to an important design requirement: a summary should not claim more authority than its sources. Fluent wording does not compensate for a missing premise or an erased disagreement. Uncertainty can be part of an accurate answer.
A correction needs to reach its consequences: scope and limits
A statement may have been used in a summary, a search result or another derived view. When the source changes, the information that depends on it needs to be reconsidered. Lineage is the relationship between information and the material from which it was derived.
That relationship helps explain why a response needs updating. It does not grant permission to keep deleted content indefinitely. Revocation and erasure are also constraints in the Lab programme: document history must not become a way around current access rights.
The scope of this requirement matters. Describing a research principle does not mean that every dependency is perfectly tracked by every connected tool. Verification should cover the transformations actually performed and allow a reader to inspect an authorized source.
For users, the intended benefit is practical. They should be able to ask what supports a response, recover a qualification and correct the underlying record when necessary. Memory becomes useful when it can be questioned, as well as consulted.
Read about decisions and their reasons and sharing context within access boundaries. Sovereignty explains the service’s broader trust boundaries.
Related reading: AI Memory Without Sources Is Just a Persistent Opinion.


