A rep removes a prospect because the company buys software through its parent. A week later, a new campaign brings the same subsidiary back. The research is accurate. The rep's decision has disappeared.
That is the memory problem worth solving in sales. A system needs to retain the reason work changed, attach it to the right part of the business, and reconsider it when the circumstances change. Saving another conversation summary does not settle any of those questions.
For Fama, company memory is part of a larger ambition: a sales function that improves as the organization learns. Fama is still in development. The design principles below describe the behavior that ambition requires; they are not a claim that a finished product already delivers it.
A Correction Has a Scope
Consider a fictional seller of maintenance scheduling software. It serves businesses whose local operations teams choose their own systems. During research, a rep rejects a regional operator because its parent controls procurement.
The instruction “exclude this company” leaves several decisions unresolved. Does it apply to this legal entity, every subsidiary, this campaign, or every future offer? Is the reason a permanent restriction, a current purchasing arrangement, or the rep's uncertainty?
The statement worth preserving is narrower: For this offer, hold the subsidiary out of campaigns aimed at local purchasing teams. The parent appears to own procurement. Reconsider if purchasing responsibility changes. The supporting source and the person who made the decision travel with that statement.
Now another rep can distinguish an intentional exclusion from an incomplete record. They can also approach the parent for an appropriate reason without treating the earlier correction as a ban on an entire corporate group.
Scope is what makes a correction reusable without making it overbroad.
Give Different Information Different Authority
A sales organization accumulates facts, preferences, policies and guesses. They should not enter memory with the same authority.
| Information | Appropriate Scope | What Can Change It |
|---|---|---|
| Approved product capability | Offer and product version | A current approved product decision |
| Account purchasing arrangement | Account, offer and time | Better evidence or a confirmed change |
| Rep's preferred opening style | Individual sender | That sender's instruction |
| Suspected reason a deal stalled | Particular opportunity | Further evidence and review |
A new transcript is not automatically more authoritative than an approved policy. A buyer asking for an unsupported integration does not make the integration available. A rep deleting a sentence from one email does not necessarily establish a company-wide writing rule.
This is why a single “latest information wins” rule is inadequate. Recency helps determine whether a fact is current. Authority determines whether a source can establish it in the first place.
Monaco's memory design separates organization context from user context. For each sales decision, preserve which offer, account, task and time a statement governs, and who can change it.
Keep the Observation Separate from the Rule
Suppose three buyers ask whether the scheduling product can respect union shift rules. There are at least three different things to retain.
The observations are that those buyers asked. The hypothesis is that shift constraints may be important in that segment. A possible action is to investigate the requirement before approaching similar accounts.
Collapsing these into “this segment requires support for union shift rules” creates a rule the evidence has not earned. It may cause the system to exclude good accounts, overstate demand to product teams, or tell a buyer that the company understands a requirement it has barely investigated.
Keep the hypothesis reviewable. Attach the conversations that support it, record counterexamples, and identify what would make the team accept or reject it. That intermediate state is useful knowledge. It does not need to masquerade as a settled fact to be valuable.
Withdrawal Is Part of Remembering
Fama's development research implementation has a narrower form of memory: shared facts about researched entities. The inspected code rebuilds those facts from their contributing evidence and withholds disputed values. A row appearing in a result and a fact being eligible for reuse are separate decisions.
That mechanism matters because a questionable value can otherwise spread through later work. If a company-to-person relationship becomes disputed, simply adding a warning to the original row leaves every copied brief exposed to the old assumption.
This is an implementation observation, not a claim of complete verification or deployed conversation memory. It does, however, establish a design commitment worth explaining: reusable information needs a path out when its support weakens.
The same principle applies to commercial context. If a product capability is withdrawn, identify the drafts and recommendations that relied on it. A new correct answer is useful; a queue of old incorrect drafts still needs attention. Rechecking those dependencies is part of the correction.
Retrieve What the Decision Needs
For the maintenance seller, a first-contact draft may need the approved offer, the account's operating model, the purchasing uncertainty and any earlier conversation. It probably does not need every lost deal, every positioning experiment, and every stylistic preference in the organization.
Anthropic's context engineering guidance treats context as something to curate for the task. Applied here, the question is straightforward: which pieces of memory could change this decision?
There is also a boundary between information that helps a rep think and information suitable for a buyer. An internal note about negotiating room can inform preparation without becoming a sentence in an email. Retrieval and permission to disclose must be separate choices.
Test the Next Task
To evaluate memory, begin with an account the system selected. Correct the purchasing arrangement, then start a new task with the same account under a slightly different name. Inspect whether the correction still applies and whether the explanation preserves its original scope.
Next, introduce evidence that local purchasing authority has returned. A useful system should help reconsider the exclusion. It should not cling to an old instruction merely because it successfully remembered it.
The final test is a different offer. If the parent-procurement restriction only mattered to the original product, the system should not silently generalize it to everything the company sells.
A sales memory earns trust when it preserves the team's reasoning without freezing the team's judgment. The next rep should inherit the lesson and the means to question it.