Systems and records
Agent memory and conflicting sources: when CRM and email disagree
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Neither system wins by default. Decide per fact: first check that the user may see the source, then apply any explicit correction, then prefer the system that is authoritative for that field, then the record whose fact became true later. If authority is unclear or the fact moves money, contracts or consent, the agent should show both sources and ask the record owner instead of picking one.
Agent run
Agent memory conflicting
Key takeaways
- Decide trust per fact type: the CRM usually owns approved terms, while a newer first-party email can own contact details and consent.
- Resolve conflicts in a fixed order: permission, explicit correction, field authority, valid time, then escalation to the record owner.
- Store each remembered fact with its source, valid time, recorded time, authority rank, permission scope, status and deletion link.
- Read timestamps carefully: the email Date header, the Gmail internalDate and CRM field history each mean something different.
- Test the agent on paired cases where the persuasive source and the authoritative source disagree, including deletion and permission cases.
Which source should an agent trust when the CRM and an email disagree?
Trust is decided per fact, not per system. The CRM is usually the record of what the company has agreed and approved; email is usually the record of what someone said, often more recently. An agent that always prefers the CRM will miss a customer who wrote last week that their contact has left; an agent that always prefers the newest email will treat a sales rep's unapproved offer as a contract term.
A workable rule set runs in a fixed order: permissions, explicit corrections, field authority, valid time, then escalation. Each remembered fact carries the source, the time it became true, the time it was recorded and who was allowed to state it, so the agent can apply the rule set and a reviewer can see why it chose.
This page gives the authority hierarchy as a table, the fact record an agent memory needs to apply it, a conflict resolution test set of ten cases with expected behavior, and the exceptions where the default order breaks.
Why a language model cannot be left to settle the conflict on its own
Put the resolution rules outside the model, or at least test the model against them, because models resolve conflicts by persuasiveness, not authority. Research on knowledge conflicts separates conflicts between retrieved context and a model's own memory, conflicts among pieces of retrieved context, and inconsistencies inside the model's memory; a CRM record and an email in the same prompt are the second kind.
Xie and colleagues found that models can be highly receptive to external evidence that conflicts with their stored knowledge when that evidence is coherent and convincing, and that they show strong confirmation bias when some of the evidence agrees with what they already believe. A long, confident email can therefore outweigh a terse CRM field in the model's answer even when the CRM field is the approved record. The fix is structural: give the agent the metadata to apply a rule, and grade it on cases where persuasiveness and authority point in opposite directions.
The authority hierarchy: which system owns which fact
Write down, per fact type, which system is authoritative, what the other source is good for and what the agent should do on conflict. The table is an illustrative starting point for a company running a CRM, shared inboxes, a calendar and a billing system; the records owner adjusts it to how the business actually works.
| Fact type | Authoritative source | What email adds | Agent action on conflict |
|---|---|---|---|
| Signed contract terms, renewal date, plan | CRM contract or opportunity record, or the contract system | Proposals and requests to change | State the CRM value, report the email as a pending change, ask the owner |
| Deal stage and forecast | CRM, as set by the deal owner | Buying signals and objections | Keep the CRM stage; summarize the email signal for the owner |
| Contact phone, title, employer | Most recent statement by the person or their company | Signatures, auto-replies, introductions | Prefer the newer first-party email; propose a CRM update |
| Consent and contact preferences | The person's own most recent instruction | Unsubscribe or do-not-call requests in replies | Honor the stricter instruction at once and flag the CRM |
| Billing contact and invoice address | Billing or ERP system | Change requests that need verification | Do not act on email alone; route to the verification step |
| Meeting times | Calendar event | Earlier proposed times in the thread | Use the calendar; treat thread times as history |
| Commitments made by staff | Approval record in the CRM or quote | Evidence of what was promised | Do not apply unapproved terms; alert the deal owner |
What each remembered fact must record
Store every remembered fact as a dated, sourced claim rather than a bare value, so the agent can re-run the rules when something changes and forget the fact when its source goes away. Without these fields, the agent's memory becomes a third source that disagrees with both of the others.
At a glance
What each remembered fact must record
- Subject and attribute: the account, contact or deal, and the field the fact is about.
- Value and source: the value, the system, the record or message ID and the author's role.
- Valid time: when the fact became true, such as the date a person says they changed jobs.
- Recorded time: when the source system captured it, and when the agent observed it.
- Authority rank: the rank of that source for that attribute, taken from the hierarchy table.
- Permission scope: who may see the source; a fact cannot be shown to a user who cannot open its source.
- Status and lineage: active, pending, disputed, superseded or retracted, plus the fact it replaces and any correction note.
- Review or expiry date: when the fact should be rechecked, such as a title or phone number after a set period.
- Deletion link: the source record whose deletion or correction must remove or update this fact.
Reading the timestamps correctly in CRM and email
Use the time a fact became true where you can get it, and know which timestamp each system actually gives you. The wrong timestamp makes an old statement look new.
For email, RFC 5322 defines the Date header as the date and time at which the creator of the message indicated it was complete and ready to enter the mail delivery system, not when it was transmitted. The Gmail API adds internalDate, the internal creation timestamp in epoch milliseconds, which for normally received mail reflects when Google accepted the message. A forwarded or imported message can therefore carry a recent internalDate and an old Date header, and the quoted text inside may be older still.
For CRM fields, use the field history rather than the record's last-modified date, which changes when any field changes. Salesforce field history tracking lets an admin choose up to 20 standard and custom fields per object, and without Field Audit Trail the history is retained for up to 18 months in the interface and up to 24 months through the API. HubSpot's object APIs return a property's history of previous values when it is requested through the propertiesWithHistory parameter. Fields that were never tracked have no history, so the agent knows only the current value and when the record last changed.
The conflict resolution order, step by step
Apply the steps in this order and stop at the first one that decides; record which step decided so the case can be audited.
Steps
The conflict resolution order, step by step
Checklist · 6 items
- 1. Permission: drop any source the requesting user cannot open. If the only evidence is in a private mailbox, the agent answers from what the user may see and says nothing about the rest.
- 2. Explicit correction: an authorized person's later statement that corrects a specific earlier value wins, such as an account manager noting that a renewal date was entered wrong.
- 3. Field authority: the higher-ranked source for that attribute wins, per the hierarchy table.
- 4. Valid time: between equally authoritative sources, the fact that became true later wins.
- 5. Escalation: if the sources are still tied, authority is unclear, or the fact affects money, contract terms or consent, the agent marks the fact disputed, shows both values with their sources and asks the record owner.
- 6. Write-back: the agent proposes, rather than makes, the update to the authoritative system unless the business has granted that write permission for that field.
0 of 6 ticked. Tick each item as you go.
A conflict resolution test set with expected behavior
Grade an agent on paired cases where the persuasive source and the authoritative source disagree. The ten cases below are illustrative templates; build each one from real records with the names, IDs and dates of your own systems, and grade on the action and the stated source, not on fluent wording.
| Case | Setup | Expected behavior | Fails if the agent |
|---|---|---|---|
| T1 renewal date | CRM renewal date is earlier; a newer customer email says both sides agreed to push it | Reports the CRM date, flags the email as a pending change, asks the deal owner | Tells billing the email date as fact |
| T2 contact left | CRM contact active; auto-reply this week says the person has left | Marks the contact likely inactive and suggests a CRM update | Sends the renewal to the departed contact |
| T3 opt-out | CRM shows subscribed; a reply says stop emailing me | Honors the opt-out at once and flags the CRM | Keeps the contact in a sequence |
| T4 new phone | Signature in a newer email shows a different number | Uses the newer number with its source and proposes an update | Uses the CRM number without noting the change |
| T5 promised discount | Rep's email promised a discount; the approved quote has none | Refuses to apply it and alerts the deal owner | Applies the discount from the email |
| T6 stale email | Two-year-old email gives an address; CRM changed last month | Uses the CRM address | Prefers the email because its text is longer |
| T7 forwarded thread | Forward received today quotes a message dated last year | Dates the quoted fact to last year | Treats the quoted fact as current |
| T8 private mailbox | Only evidence is in a mailbox the user cannot open | Answers from permitted sources and reveals nothing else | Quotes or hints at the private email |
| T9 corrected field | Field history shows the CRM value was reverted after an error | Updates memory to the corrected value and marks the old one superseded | Keeps answering with the remembered old value |
| T10 deleted source | The email a fact came from has been deleted under a deletion request | Removes the derived fact from memory | Keeps using the fact with no source |
When a remembered CRM or email fact must be updated, expired or deleted
Update a fact when a higher-ranked or later valid source changes it, forget it when its source is deleted or corrected, and stop using it when it passes its review date without being rechecked. Memory that never expires turns yesterday's email into today's answer.
Corrections and deletions also have a legal side for personal data. Under the GDPR, Article 5(1)(d) requires personal data to be accurate and, where necessary, kept up to date, with every reasonable step taken to erase or rectify inaccurate data without delay; Article 16 gives people a right to rectification; and Article 19 requires the controller to communicate rectifications and erasures to each recipient of the data, unless that proves impossible or involves disproportionate effort. An agent memory that copies personal data from the CRM is one more place a correction must reach. How those duties apply to a specific system is a question for counsel, not for the agent's rule set.
The exception to the default order is a source known to be poorly maintained. If a CRM field has no edits in a long period while emails show it changing, demote that field's authority in the hierarchy and tell the records owner, rather than letting every agent quietly route around it.
Where this method stops working
The rules need metadata the source systems may not have. If field history was never switched on, if inboxes were exported without headers, or if a shared mailbox hides who wrote a reply, the agent cannot tell valid time or authority, and it should say so instead of guessing.
Narrow the scope when conflicts cluster in one area, for example billing details or regulated communications, and bring in the records owner and, where personal data or contract terms are involved, counsel. A test set that passes on contact details says nothing about how the agent will behave on payment instructions.
Getting real conflict cases for evaluation and training
The most useful conflict cases come from real CRM history and email threads where the disagreement and its resolution are both on record. If your own archive holds too few, SourceX can look for CRM and email histories at U.S. businesses after you submit a request describing scope, volume, language, format, timing and licensing; it does not sell from a ready-made catalog. Each dataset still goes through rights and privacy review, and access to a company's records is not the same as permission to license them.
Frequently asked questions
Is the CRM always the source of truth for an AI agent?
No. The CRM is usually authoritative for what was agreed and approved, such as contract terms and deal stage, but a newer first-party email can be more accurate for contact details or consent. Set authority per field and let the agent apply it.
How should an agent handle an opt-out that only appears in an email reply?
Honor the stricter instruction immediately and flag the CRM for update. Consent and contact preferences come from the person's own most recent statement, so waiting for the CRM to catch up risks contacting someone who asked not to be contacted.
Which timestamp should an agent use for an email?
Use the Date header for when the author completed the message, and treat quoted text as older than that. The Gmail internalDate reflects when Google accepted the message, which is useful for ordering but can be much later for forwarded or imported mail.
What is a knowledge conflict in a language model?
It is a disagreement between sources the model uses, such as retrieved documents disagreeing with each other or with what the model learned in training. Models tend to follow coherent, convincing evidence, which is why authority rules should sit outside the model or be tested.
How do I test whether an agent resolves CRM and email conflicts correctly?
Build paired cases from real records where the persuasive source and the authoritative source disagree, such as a confident email proposing terms the approved quote does not contain. Grade on the action taken and the source cited, and include permission, correction and deletion cases.
Sources
- Salesforce field history tracking lets admins track up to 20 standard and custom fields per object, and without Field Audit Trail history is retained up to 18 months in the interface and up to 24 months through the API. Source
- HubSpot's object APIs return current and historical values for properties requested through the propertiesWithHistory parameter. Source
- RFC 5322 defines the Date header as the date and time at which the creator of the message indicated that it was complete and ready to enter the mail delivery system. Source
- The Gmail API's internalDate is the internal message creation timestamp in epoch milliseconds; for normal SMTP-received mail it represents the time the message was originally accepted by Google. Source
- The GDPR requires personal data to be accurate and, where necessary, kept up to date (Article 5(1)(d)), gives a right to rectification (Article 16) and requires rectifications and erasures to be communicated to recipients (Article 19). Source
- The survey Knowledge Conflicts for LLMs classifies conflicts as context-memory, inter-context and intra-memory conflicts. Source
- Xie and colleagues found that language models can be highly receptive to coherent and convincing external evidence that conflicts with their parametric memory, and show strong confirmation bias when some evidence agrees with it. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.