Open Data Infrastructure
Data Modeling for Temporal Entity Memory
How entity memory needs valid time, observed time, source confidence, consent state, and decay rules before agents trust history.
Agent memory gets dangerous when history flattens into "known facts."
Entity memory needs more than a latest row
An entity memory model stores what the system believes about a person, account, device, product, customer, supplier, or process. For agents, that memory needs time. The model should separate valid time, when a fact was true in the world, from observed time, when the system learned or recorded it.
Without that split, agents confuse old truth, late-arriving truth, corrected truth, and revoked truth. That is how a helpful assistant becomes confidently wrong.
Source confidence and consent are part of the model
Temporal entity memory also needs source confidence, source authority, consent state, policy constraints, and decay rules. An agent should know whether a value came from a system of record, an inferred signal, a customer statement, or a stale derived table.
Provenance standards and lineage systems help because memory needs evidence. A value without provenance may still be useful for search. It should not automatically become trusted operational context.
Core idea: entity memory is a temporal contract, not a bag of attributes.
The ODI pattern ties memory back to systems of record
Open Data Infrastructure connects entity memory to source systems, lineage, policy, freshness, and ownership. That connection matters because agents use memory to decide what to retrieve, what to ask, and what to do next.
For related patterns, see consent-aware data modeling, identity resolution for agentic systems, and AI-ready context decay budgets. Memory is useful only when its trust boundary is modeled.
What breaks first
Temporal memory failures often look like reasoning failures. The model is reasoning over a bad time model.
- A current profile overwrites historical facts needed for audit.
- A late-arriving correction is treated as if it was known earlier.
- Consent status changes, but derived memory remains available.
- An inferred attribute is ranked like a system-of-record value.
Questions to ask while modeling
Ask when the fact was true, when the system observed it, who or what asserted it, how confident the system is, which policy applies, and when the memory should decay.
A temporal entity model should make old context useful without making old context pretend to be current.
Sources to start with
These primary sources anchor the technical claims in this guide.
- W3C PROV overview
- OpenLineage object model documentation
- DataHub lineage documentation
- OpenAI evals guide
Agent memory should remember history without erasing time.