Sales and Support Agents: Share Verified Facts, Keep Private Notes
How to split sales and support agent memory so only owner-verified facts cross between agents and private notes stay in their lane, with a practical plan.

Sales and support agents can share memory safely when only verified, owner-approved facts move between them and every working note stays with the agent that wrote it. The method is to keep three layers apart: private working notes for each agent, a small shared store of confirmed facts that each carries a named owner and a source, and a rule that a note becomes shared only after a person approves it. Done this way, a sales guess never becomes a support promise.
Key takeaways
- Give each agent a private layer for drafts, impressions and intermediate steps that never leaves its lane.
- Let only confirmed facts, each with an owner and a source, live in the shared layer.
- Make promotion from private to shared a reviewed step, never a side effect of a conversation.
- Scope retrieval so each agent sees the shared facts its role needs, and decide the overlap on purpose.
- Log every promotion, correction and withdrawal so a wrong fact can be traced and removed.
Why sales and support need different memory, and where they overlap
A sales agent accumulates deal context: what a buyer said matters most to them, which competitors came up, how a discount conversation went, and what the account team expects to close. A support agent accumulates service context: the ticket history, the steps already tried, the error that was reproduced, the workaround that partly worked and the customer's frustration. Most of this material is useful only inside its own function, and much of it is opinion. Writing it into one shared pool would make every later answer harder to trust.
The overlap is smaller but more dangerous. Both agents need the current plan names, the contract terms that apply to an account, the published policy version, the named account owner, communication preferences and any open commitment that the company has actually made. These are the facts where a mismatch causes damage. If sales says a feature will be available soon and support repeats it as a promise, the customer will hold the company to that statement.
Illustrative scenario: a sales agent records that a prospect was told an integration would be ready soon. In a shared pool, a support agent later reads that line as a commitment and tells the customer the same thing. In a split design, the sales note stays private, while the shared store holds only the product status that the product owner has confirmed, which says the integration is not committed. The customer hears the confirmed status, and the internal conversation remains internal.
Sort every memory item into one of three lanes
Before any agent writes to shared memory, each item needs a lane. A simple decision guide works for most teams:
- Private working note: a guess, draft, impression, opinion, or intermediate step. It stays with the agent that wrote it and is never retrieved by the other agent.
- Shared verified fact: a statement that a named owner has checked against a source of truth, such as a current plan name, a contract clause, or the published version of a policy.
- Restricted record: personal data, internal pricing decisions, or financial terms that only specific roles may read. Sensitive personal details should usually not be stored at all, which is covered in sensitive data in AI agent memory.
Two questions settle most cases. First, would another agent act on this without asking anyone? If yes, it needs verification before it is shared. Second, would a wrong version cause a customer, a contract or a person harm? If yes, it needs a named owner and a restricted scope. Anything that describes a person's view of a customer, such as a judgment about budget or tone, belongs in the private lane by default.
Build the promotion path, step by step
Promotion is the only door between lanes, so it should be designed as a short, explicit process:
1. Name an owner for each fact category. Sales operations can own deal terms, the support lead can own service policy, and the product owner can own capability statements. 2. Define the shape of a shared fact: the statement itself, the owner, the source document or system, the version, the roles allowed to read it, and a status of confirmed, disputed or withdrawn. 3. Let an agent propose a promotion by writing a candidate. A candidate is never treated as a confirmed fact. 4. The owner checks the candidate against the source, then accepts it, edits it or rejects it. 5. An accepted fact enters the shared store with its source attached, and both agents retrieve it with a citation. 6. When someone disputes a fact, the owner marks it disputed, both agents stop stating it as settled, and the change is logged. 7. When the source changes, the owner reviews the fact again or withdraws it.
The rules for who may write and how changes are recorded are explained in shared memory write rules for teams, and the gate itself is described in memory write gates for superintelligent agents. Keep the reviewer list short. A shared store that a dozen people must approve will quickly fill with unreviewed notes that everyone assumes are checked.
Control what each agent can retrieve
Separating lanes on paper is not enough. The retrieval layer must enforce the same boundary. Access decisions should be made by the system from the agent's role and the scope of each record, not by instructions inside the prompt or by something a customer typed. The OWASP guidance on authorization makes the same point: checks belong in the enforcing layer, and they should deny by default when a role or scope is missing, as described in the [OWASP authorization cheat sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html).
In practice, the support agent can read the shared product status, the policy versions and the account service history. The sales agent can read the shared plan names, the account owner and the confirmed contract terms. Neither agent can read the other's private layer. Where the two roles genuinely need the same record, the overlap is written down as a rule, so nobody discovers it by accident. For a fuller treatment of how departments see different records, see memory scope by department.
Retrieval also decides what gets cited. Every answer built from shared memory should show the source and version of the fact it used. If the answer cannot point to a confirmed record, the agent should say so and route the question to the owner rather than fill the gap from a private note.
Keep private notes in their lane
Private notes are valuable, and the goal is not to erase them. Sales needs its meeting context, and support needs its diagnostic trail. The discipline is that these notes stay useful to their author and do not leak into answers for other roles. Three habits help. Agents should never quote another agent's private note, even when it is retrieved by mistake, because a retrieval error should be reported as an incident. Summaries that move a note toward the shared store must go through promotion, not through a silent copy. And context compaction should keep decisions and open commitments while dropping speculation, a point that is covered in AI context compaction without losing important decisions.
Review private layers on a schedule. Old drafts that were never promoted should expire, and any note that contains a commitment to a customer should be escalated to the owner for a decision. This keeps the private layer from becoming a second, unreviewed company memory.
How Be Human handles the split in managed operations
Be Human designs the lanes during a first engagement, defines the owners and the fact schema with your team, and implements the retrieval scopes and the promotion flow so that the rules are enforced by the system. In managed operations, Be Human monitors whether agents cite confirmed facts, runs recurring evaluations on boundary cases, reviews permissions when roles change, and handles incidents such as a private note appearing in a shared answer. Improvements to the split are made in controlled steps with an owner's approval.
Test that the split holds
A split is only real when it survives tests. The memory test cases for superintelligent agents describe how to check recall and forgetting, and the same approach applies here. Four tests cover most of the risk. Ask the sales agent about a support-only detail and confirm it cannot retrieve it. Ask the support agent about a discount discussion and confirm it does not reveal one. Submit a candidate fact with a false claim and confirm it stays out of the shared store until an owner accepts it. Withdraw a confirmed fact and confirm that both agents stop stating it.
The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) is a useful reference for naming owners, documenting controls and deciding how often they are measured. If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Can sales and support agents read each other's notes? Not by default. Each agent has a private layer that only its author can retrieve, and the shared layer holds only confirmed facts. Any exception should be written as an explicit rule with an owner, so that it is visible and reviewed.
Who should approve a fact before it becomes shared? The person who owns the source of that fact should approve it. Sales operations approves deal terms, the support lead approves service policy, and the product owner approves capability statements. The agent may propose the fact, but it should never confirm its own proposal.
What should a shared fact record contain? It should contain the statement, the owner, the source document or system, the version, the roles allowed to read it, and a status such as confirmed, disputed or withdrawn. Without the source and version, nobody can check the fact later, and the agent cannot cite it in a way a staff member can verify.
What happens when a shared fact turns out to be wrong? The owner marks it disputed or withdraws it, and both agents stop presenting it as settled. The change is logged with the reason, so the team can see which answers may have used the wrong fact and correct them where customers received it.
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.


