Article chapter 03 of 08
Give every record provenance and validity
From Long-term AI memory: records, relationships and retrieval
A durable memory record needs fields that identify what it means, where it came from, when it applied and whether it can still be used. Text and an embedding do not carry that structure by themselves.
A record can include:
memory_idand a stable type;- a structured subject and property where appropriate;
- the value or concise statement;
- source object and source location;
- source author or system identity;
- recorded time and effective time;
- validity status and optional expiry;
- confidence or extraction status;
- sensitivity and access scope;
- the record that superseded it;
- the rule or process that created it.
Use effective time separately from recorded time. A document can be uploaded after the decision it describes took effect. A future roster change can be recorded in advance but should not be returned as current until its start date.
Provenance should be precise enough to reopen the evidence. A document ID without a page, heading or paragraph location makes review slow. A transcript reference should identify the message or bounded span. If the source can change, retain a version or checksum so the remembered claim remains tied to the material actually reviewed.
Assign authority by source and record type, not by confidence language generated by a model. An approved project decision may outrank an informal chat suggestion. A system-of-record field may outrank a summary copied into meeting notes. Encode those rules so retrieval can resolve conflict consistently.
Memory text can still carry uncertainty. "The preferred date may be Friday" should not be normalised into a confirmed Friday deadline. Store an assertion class such as confirmed, proposed, inferred or disputed, then make the retrieval interface expose it.
Sensitive records need field-level thought. A broad project scope should not automatically grant access to personal notes linked to someone who works on that project. Apply the requesting user's access before ranking, and avoid putting restricted values into search indexes that a broader service can inspect.