One question can require several records
Consider a routine business question: Does the pricing exception approved last quarter still apply to this contract?
No single document necessarily contains the answer.
| Record | What it contributes |
|---|---|
| Meeting notes | An exception was approved for a particular customer situation. |
| Contract | The agreement defines the account and conditions to which the exception may apply. |
| Later policy notice | Approval authority changed after the meeting, so the exception’s current validity must be reviewed. |
The records do not merely add three passages to a prompt. They play different roles. One records a decision, another constrains its scope, and a third changes the conditions under which someone should rely on it.
Provenance keeps the reasoning inspectable
A retrieved statement is more useful when a person can see where it came from and what other records qualify it. Source location, document identity, date, revision information, and the relationship proposed between records can all help a reviewer reconstruct why a passage was selected.
Consilience is designed to keep original material available while AI-assisted processing proposes entities, aliases, and relationships. Those proposals are not self-authenticating facts. Knowledge owners need to be able to confirm, correct, split, merge, or reject them.
Time is a signal, not an automatic verdict
A newer document does not always replace an older one. A policy may apply only to new contracts, an exception may survive a general revision, and an effective date may differ from a publication date. Dates and revision signals help narrow the review; they do not by themselves establish legal or operational validity.
For this reason, EngramRAG treats temporal and provenance information as part of the context made available for review. The system can surface a chain from decision to scope to later change. A person or an explicitly defined workflow still determines what remains valid for the decision at hand.
Similarity alone does not establish the chain
Semantic and lexical retrieval can locate relevant passages. The remaining task is to understand how those passages relate: who decided what, for whom, under which conditions, at what time, and based on which evidence. A memory architecture should make those relationships visible without pretending that every ambiguity has already been resolved.