Skip to content

Systems and records

Enterprise context provenance scoring: give agent memory a source, date and confidence

By SourceX Editorial · Updated

10 sections

Short answer

Store each remembered fact as a claim, not a sentence: the source record it came from and how it was derived, when it became true and when it stops being true, when the agent learned it, who may see it, and a confidence band built from source authority, extraction method, corroboration and freshness. The band decides whether the agent uses the fact silently, cites it, confirms it first or leaves it out.

Provenance trail

Enterprise context

This page covers enterprise context provenance scoring: give agent memory a source, date and confidence.Illustrative

Key takeaways

  • Store each memory as a claim with a resolvable source, a derivation type, validity dates, a record time and an access scope.
  • Storage timestamps such as created_at say when the agent wrote a fact, not when it was true.
  • Score provenance with permission and status vetoes first, then ordinal bands from authority, derivation, corroboration and freshness.
  • Each band must map to one agent behavior: use, cite, confirm or withhold.
  • Consent, commitments, prices and access rights are read from the system of record at use time, never from memory alone.

What source, date and confidence metadata should accompany agent memory?

Every memory item needs four groups of metadata. Source: a resolvable pointer to the record or message it came from and how the fact was derived from it. Time: when the fact became true, when it stopped being true, and when the agent recorded it. Permission: the access scope of the source, carried over to the memory. Confidence: a band computed from those fields, not a number the model guesses about itself.

With those fields an agent can answer the questions a business user will eventually ask about any remembered fact: where did this come from, is it still true, am I allowed to see it, and how sure should I be. Without them, the agent holds a summary it cannot check, cannot correct and cannot safely forget.

This page gives a field-level memory item contract, a provenance scoring method with veto rules, the behavior each confidence band should trigger, an illustrative worked example and the facts that should never be settled by a score.

Why can't a summary without provenance be checked or corrected?

A memory such as Acme renews in March, prefers invoicing by email reads as useful and is impossible to audit. It does not say whether March came from the signed contract, a sales rep's call note or the model's reading of a forwarded email. When the contract is amended, nothing links the memory to the change, so the stale date keeps coming back with the same confidence as the day it was written.

Provenance turns that sentence into claims that can be maintained. The W3C PROV ontology gives a vocabulary that fits: an entity can be attributed to an agent, generated at a time and derived from another entity, with more specific derivation types for quotation, revision and primary source. A memory store does not need to adopt PROV wholesale, but those distinctions are the ones the contract below keeps.

What do current memory stores record, and what gaps do they leave?

Memory frameworks record less than a business needs by default, and what they do record is easy to misread. LangGraph's base store keeps each item as a value dictionary under a namespace and key, with created_at and updated_at timestamps. Those are storage times: they say when the agent wrote the item, not when the fact was true. Its optional time-to-live expires items a set number of minutes after last access, and reads refresh the timer by default, so a frequently used stale fact lives longest.

Graphiti, an open-source temporal knowledge graph framework, goes further. Its README says facts carry validity windows and that old facts are invalidated rather than deleted when information changes, with every derived fact traced back to the episodes, the raw data, that produced it. Its edge model separates valid_at, when the fact became true, from invalid_at, when it stopped being true, and from expired_at, when the edge was invalidated in the graph, and it keeps a reference time from the source episode.

Neither design records the permission scope of the source, the authority of the system it came from, or how the fact was extracted. Those are the fields an enterprise has to add, and they are the inputs a provenance score needs.

Which fields belong in the memory item metadata contract?

Treat the table as a schema. Required fields must be present before an item can be retrieved for any user; recommended fields can be empty but must be explicitly null, so a missing value is never mistaken for a default.

  • Keep an append-only correction log per item: who or what changed it, when, and the new source_ref.
  • Record the extractor version or prompt version for model_extracted items, so a bad extractor can be traced and its items demoted in bulk.
  • Store a retention rule reference rather than a free-text note, so deletion follows the source's policy.
Which fields belong in the memory item metadata contract?
FieldMeaningIllustrative valueRequired
claimOne fact in subject, attribute, value formAccount 4471, renewal_date, 2027-03-31Yes
source_refResolvable pointer to the record, field or messagecrm:opportunity/0068x/field/CloseDateYes
source_system_roleWhether the source is the system of record for this attributesystem_of_record, secondary or conversationalYes
derivationHow the value was obtained from the sourcecopied_field, quoted_text, model_extracted or inferredYes
valid_from and valid_toWhen the fact became true and stopped being true, if known2026-04-01 to nullvalid_from yes
observed_atSource timestamp of the value: field history entry or message sent time2026-04-03T15:12ZYes
recorded_atWhen the agent wrote the memory item2026-04-03T15:20ZYes
access_scopePermission scope inherited from the sourcegroup:finance-ar plus account ownerYes
supersedes and corroborated_byItems this one replaces and independent sources that agreemem_9921; email msg 1834Recommended
status and confidence_bandActive, superseded, retracted or expired; band from the scoring rulesactive, BYes

How should source timestamps in CRM records be read?

The observed_at field is only as good as the source's own history, and CRM history is often partial. Salesforce field history tracking records old and new values only for fields an administrator selected, up to 20 per object in the standard feature, and starts from the moment tracking is turned on; for long text fields it records that a change happened without storing the values. A renewal date on a field that was never tracked has no recoverable change date.

HubSpot shows a property's history on the record with the value, the date and the source of each change, such as an import, an integration or a workflow, and lets an administrator export that history. That change source is useful input for the derivation and authority fields: a value set by a sales rep in the record carries different weight from one written by an enrichment integration.

When the source has no field history, set observed_at to the last modified time of the record, mark the item's time basis as record_level, and let the scoring rules treat it as weaker evidence. Never copy recorded_at into observed_at to fill the gap.

How do you score provenance without inventing a probability?

Score with ordinal bands and explicit vetoes, not a blended decimal. A number such as 0.82 invites the agent and its users to read it as a calibrated probability, which it is not unless you have measured it against verified outcomes. Bands built from named components can be explained in one sentence and audited by a records owner.

Apply the vetoes first. An item whose access_scope excludes the current user is not retrieved at all. An item with status superseded, retracted or expired is not used as current, whatever its other scores. Only then combine the four components below into a band.

Permission scope needs one rule of its own. When a memory item is derived from several sources, such as a call note and a contract, its access_scope is the intersection of their scopes, not the union, so a fact that needed the contract to be confirmed stays invisible to anyone who cannot open the contract. Re-evaluate the scope when a source's permissions change, not only when its content changes.

Bands can become calibrated later. Once enough items have been checked against their source at use time, count how often each band was right and adjust the rules, for example by moving model-extracted values from a commonly mis-read field down a band. Keep the rule set versioned in the store, so every item records which version graded it and a change of rules can be replayed over existing memory.

How do you score provenance without inventing a probability?
ComponentStrongMediumWeak
Source authoritySystem of record for this attributeSecondary system or signed document copyConversation, note or third-party enrichment
DerivationCopied field valueQuoted text from a dated messageModel-extracted or inferred
CorroborationTwo independent sources agreeOne source, no conflict foundA live conflict exists
Freshness against validityInside a known validity windowNo window; observed recently for its attribute typeOlder than the attribute's review interval

What the agent does at each confidence band

The band only matters if it changes behavior. Map each band to one action and test that the agent follows it; a contract that the agent ignores at answer time is documentation, not control.

What the agent does at each confidence band
BandRuleAgent behaviorWhat the user sees
AAll four components strong, or three strong and none weakUse the fact directlyAnswer, with source available on request
BNo weak componentUse the fact and cite itAnswer with source and observed date
COne weak componentRe-read the source before use, or confirm with the ownerA check step, then the answer
DTwo or more weak components, or a live conflictDo not assert; show what is known and from whereBoth values, sources and dates, and a question

How should a renewal date remembered from a call note evolve? (illustrative)

This example is illustrative; the account, dates and records are invented. In April a sales rep's call note says the customer expects to renew at the end of March next year. The agent extracts renewal_date 2027-03-31 with derivation model_extracted from a conversational source, observed on the note's timestamp. No system of record agrees yet. Authority weak, derivation weak: band D, so when finance asks, the agent reports that a call note suggests end of March and that the contract record has no renewal date.

In May the signed order form is attached to the opportunity and the CRM close date field is set to 2027-03-31, with field history tracking on. The agent writes a new item from the system of record, derivation copied_field, corroborated by the call note. It is band A, and the old item gets status superseded with a pointer to the new one.

In September an amendment moves the renewal to June 2027 and the field changes. Field history gives a new observed_at; the March item receives valid_to and status superseded, and the June item becomes current. The agent can now answer both when does Acme renew and what did we believe in May, which is the question auditors and account managers ask after a missed renewal.

When must a remembered fact be refreshed, demoted or set aside?

Some facts should never be settled by a score. Consent and communication preferences, signed commitments, prices and payment terms, identity and access rights must be read from their system of record at the time of use, whatever band the memory holds; the memory may point to the record but should not stand in for it. Health, legal and employment details call for the source's own retention and access rules, and a records owner should decide whether they enter memory at all.

At a glance

When must a remembered fact be refreshed, demoted or set aside

For everything else, write the triggers down so they run without judgment calls.

  • Refresh when the source record changes: subscribe to field history or change events where the source offers them.
  • Demote one band when an item passes the review interval for its attribute type without a new observation.
  • Set status superseded when a higher-authority source gives a different value, and keep both items linked.
  • Set status retracted when the source itself is corrected or withdrawn, and suppress retrieval immediately.
  • Follow the source's deletion: when a record or message is deleted under policy, delete or tombstone the derived items too.
  • Stop and escalate to the records owner when two system-of-record sources disagree; the contract cannot resolve that.

Where do records that carry their own change history come from?

Building and testing a provenance score needs records whose history is intact: field-level change logs, message timestamps, amendment chains and the links between a note and the record it later updated. Exports that flatten a CRM to current values remove exactly the evidence the score depends on.

When change history matters, SourceX looks for records at U.S. businesses on a team's request rather than drawing on inventory, and the request-data form captures scope, volume, language, format, timing and licensing needs. Ask for field history and change-source columns in the delivery specification, since they are hard to reconstruct after export.

Frequently asked questions

What metadata should an AI agent store with each memory?

At minimum: a pointer to the source record, how the value was derived, when the fact became true and stopped being true, when the agent recorded it, the access scope inherited from the source, a status and a confidence band. Those fields let the agent cite, correct, expire and withhold a fact.

Is created_at enough to tell whether a memory is current?

No. In most memory stores created_at and updated_at are storage times, so they show when the agent wrote the item, not when the fact was true. Keep a separate valid-from and valid-to pair and a source observation time.

What is provenance scoring for enterprise context?

It is a rule-based grade of how far a remembered fact can be trusted, built from the authority of its source, how it was extracted, whether other sources agree and how fresh it is against its validity. Permission and status checks act as vetoes before any grading.

Should the confidence be a probability?

Only if you have calibrated it against verified outcomes. Otherwise ordinal bands with named components are more honest and easier for records owners to audit, and they map cleanly to agent behaviors such as cite, confirm or withhold.

What happens to a memory when the source record is corrected?

The memory item should be marked superseded or retracted, linked to the corrected source, and excluded from current answers immediately. Keeping the old item with its validity dates lets the agent explain what was believed earlier without presenting it as current.

Sources

  • The W3C PROV ontology describes entities attributed to agents, generated at a time and derived from other entities, with derivation subtypes for quotation, revision and primary source. Source
  • LangGraph's base store keeps items with a namespace, key, value, created_at and updated_at, and its optional TTL expires items after last access, refreshed on reads by default. Source
  • Graphiti's README says facts carry validity windows, old facts are invalidated rather than deleted, and every derived fact traces back to source episodes. Source
  • Graphiti's edge model separates valid_at, invalid_at and expired_at and keeps a reference time from the source episode. Source
  • Salesforce field history tracking records old and new values only for selected fields, starts when enabled, and for long text fields records only that a change occurred. Source
  • HubSpot shows each property change with its value, date and source on the record, and property history can be exported. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies