Skip to content
be-human.ai
← Back to the journal

Context & memory

Provenance Tags for AI Agent Memory: Trace Every Fact to Its Source

Add source, author and date tags to every AI agent memory fact so reviewers can check where it came from and whether it still holds.

Be Human typographic cover with the title: Provenance Tags for AI Agent Memory: Trace Every Fact to Its Source
Conceptual illustration created with AI · Be Human

Provenance tags attach a source, an author and a date to every fact an AI agent stores in memory, so the agent and its human reviewers can trace any remembered statement back to the document, message or person it came from. Without these tags, a memory reads as settled truth even when nobody can say where it started or whether it is still valid. Writing the tags at the moment a fact enters memory is the simplest way to keep that memory checkable.

Key takeaways

  • A memory fact without a source is an unverified claim, even when the agent states it with confidence.
  • Write the source, author and date when the fact enters memory, not during a later cleanup.
  • Use the tags to decide which fact wins when two memories disagree, and to show staff where an answer came from.
  • Review tags on a schedule, because a fact that was correct when written can go stale without anyone editing the memory.
  • Keep the tag fields few and fixed so a reviewer can scan them in seconds.

What a provenance tag records

A useful tag answers four questions about a fact: where it came from, who put it there, when the underlying information was written, and when it was stored. Those last two are different. A policy document may have been written months before an agent learned about it, and the gap matters when someone asks whether the answer is current. Keep these fields separate so the reviewer does not have to guess.

The source field should point to a stable reference, such as a document identifier with its version, a record ID in the system of record, or a message thread that a person can open. A phrase like the company handbook is not enough, because the handbook changes. The author field should name a person or a named integration, never the assistant itself. When the agent writes a summary of a conversation, the summary is a derived fact, and its author should say so.

The NIST AI Risk Management Framework treats documentation and transparency as part of managing AI risk, and provenance tags are a small, practical form of that documentation inside an agent's memory. They do not prove a fact is correct. They make it possible to check.

Write the tag when the fact enters memory

Tags added later tend to be guesses. The cheapest moment to record provenance is the moment a fact is written, when the source is still in front of the person or process that wrote it. A workable write rule for most teams has six parts:

  • Refuse to store a fact that has no source field. Hold it as a pending note that a reviewer must resolve.
  • Record the source as a stable reference, with the document identifier and version or the record ID.
  • Record the author as a person or a named integration that has permission to write that type of fact.
  • Record two dates: the date the source was written and the date the fact was stored.
  • Mark the fact as primary (copied from the source), derived (summarized or inferred by the agent), or confirmed (checked by a staff member).
  • Set a review date for facts that describe schedules, policies, service terms or staffing, since these change most often.

The rule about derived facts matters most. An agent that summarizes last week's conversation and stores the summary as a primary fact has turned its own interpretation into a source. Keep derived facts visible as derived, and do not let them be the only support for a commitment. If an agent keeps losing context between sessions, the cause is often in how facts are written, which is why stale context in AI agents deserves a look alongside provenance.

How the agent uses tags when it answers

Provenance has value only if the agent reads it at answer time. The agent should treat each retrieved fact as carrying its tags, and its behavior should follow a simple decision guide:

  • If the fact is primary, its source is current, and the review date has not passed, answer and name the source and its date when the answer depends on it.
  • If the fact is derived, or its source is older than the review window, answer only with a clear caveat, or ask the owner to confirm before relying on it.
  • If the fact has no source at all, do not state it as company fact. Route it to review.
  • If two tagged facts conflict, apply a written precedence rule, such as a confirmed primary fact over a derived one and a newer source over an older one, and log the conflict for an owner to resolve.

The last point deserves a written rule before the first conflict appears. Teams that skip it usually end up with an agent that picks whichever fact was retrieved first. The approach in resolving conflicting AI agent memories covers how to set that precedence. For answers that go to customers, the source citation should be specific enough that a staff member can open the same document in under a minute, which is the standard described in citing exact document versions.

One caution: a tag does not make a fact true. It makes the fact easier to challenge, and that is the whole purpose.

An illustrative scenario

The following scenario is illustrative and does not describe a real client. A field service company has an agent that answers customer questions about warranty terms. Its memory holds a fact about warranty length, tagged with the terms document it came from, the operations manager who confirmed it, and the date the document was written.

Later, the company publishes a revised terms document. The new version is written on a later date, and the operations manager confirms the changed warranty length. The memory now holds two facts about the same topic. Under the precedence rule, the confirmed fact from the newer document wins, and the older fact is marked as superseded instead of being deleted. The agent answers from the new fact and cites the current version. It also logs the change, so a reviewer can see when the answer changed and why.

Without tags, the agent would have had two confident statements with no way to choose between them, and the staff member answering the phone would have had no way to tell which one the customer had been given.

What reviewers should check

A reviewer does not need to read every memory fact. A short, repeatable check catches most provenance problems:

  • Each fact has a source reference that opens to the actual text or record.
  • The source version is the current one, or the fact is marked as superseded with a link to its replacement.
  • The author field names a person or integration that is allowed to write this type of fact.
  • Derived facts are not the only support for prices, policies, service commitments or schedules.
  • Facts past their review date have been confirmed, corrected or removed.
  • Conflicts in the log have an owner and a resolution date.

Run this check on a schedule rather than only after a complaint. The method used to audit what an agent remembers is covered in reviewing what an AI agent remembers, and it fits well with provenance checks because the tags give the reviewer something concrete to inspect.

How Be Human handles provenance in managed operations

In a Be Human engagement, provenance rules are part of the first definition of the agent: which sources may create memory, which tag fields are required, who may write and who may approve changes. Managed operations then keep those rules working over time. That includes recurring evaluations that sample memory facts and check their tags, permission reviews that confirm who can write which fact types, and incident response when a fact with a missing or wrong source reaches a customer answer. Success is defined as an observable criterion agreed at the start, such as the proportion of answered facts that cite a current source, and it is measured rather than promised. Be Human does not guarantee outcomes, but it does make provenance a routine part of operations rather than a one-time setup. If you are not sure which of your company's facts an agent should be trusted to remember, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Does a provenance tag make a memory fact true? No, a tag does not verify the content of a fact. It records where the fact came from, who put it there and when, so that a person or a later check can judge whether it is still correct. The value is that errors become traceable and fixable rather than invisible.

Should the agent store its own summaries of conversations as memory? It can, but those summaries should be tagged as derived and kept separate from primary facts. A summary is an interpretation of a conversation, not the conversation itself. Treat it as a lead to confirm before it supports any commitment, pricing, policy or schedule.

How much source detail should a tag contain? Enough for a staff member to open the exact text or record within a minute, and no more. A stable document identifier with its version and a link or record ID is usually sufficient. Long excerpts copied into tags make memory heavier and harder to review, so keep the excerpt short and point to the original.

What happens to tagged facts when a source document is replaced? The facts that depend on the old version should be marked as superseded, not silently deleted, and linked to the fact that replaces them. This keeps the history visible for audits and lets the agent answer from the current version. A scheduled review should catch facts whose source has changed, so that replacement is routine rather than an emergency.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →