Is there research specifically on the power imbalance created by a company remembering everything about a client relationship?

Yes — a 2025 conceptual paper from UC Riverside's School of Business names this precisely: "Memory Power Asymmetry" (MPA). It is defined as a structural power imbalance that arises when one party in a relationship can record, retain, retrieve and combine the shared history of that relationship far better than the other party can. The imbalance exists whether or not that advantage is ever actively misused.

Is this just another way of describing normal data privacy or surveillance concerns?

The researchers are explicit that it's related but distinct. MPA can exist even when privacy rules are fully respected and no data is ever leaked or misused. The imbalance comes purely from one side being able to reconstruct the shared past far better than the other. That shapes trust and behavior independent of whether that memory is ever abused. A concrete version of the same gap: the people you talk to who never consented .

What actually reduces this kind of imbalance, according to the research?

The paper proposes six concrete design principles. Three carry most of the weight: Making memory visible and explainable to the person it's about, not just to the company. Building in deliberate forgetting or retention limits rather than keeping everything indefinitely by default. Framing AI memory as something that helps the client too, rather than only as a tool the business uses on the client. Closer does not meet the second one by default: nothing expires on a timer unless an operator turns retention on, and even then it does not purge the vector store. See can you actually delete something from the memory .

Can the client see what Closer remembers about them?

Not directly. There is no client-facing login that lets the person being remembered inspect their own record. Closer shows the stored memory and knowledge graph to the account holder — the salesperson and their team. A client's memories can be exported as a markdown file, which someone on the account can share with the client. The asymmetry described in this post is reduced by that export, not removed by it. Nor can a single remembered fact be corrected in place: what if the AI remembers something wrong .

When One Side Remembers Everything: The Power Imbalance Nobody Names

Is this a real, named concept, or a general worry dressed up in academic language?

It is a specific concept, introduced in a 2025 paper by researchers at UC Riverside's School of Business. They define Memory Power Asymmetry (MPA) as "a structural power imbalance that arises when one relationship partner... possesses a substantially superior capacity to record, retain, retrieve, and integrate the shared history of the relationship, and can selectively deploy that history in ways the other partner... cannot."

The important word is structural. The imbalance exists by virtue of the technical setup, independent of whether anyone exploits it. As the authors put it: "even a benevolent, privacy-respecting AI that never abuses its memory advantage still participates in an asymmetric configuration."

One honest caveat on the source: this is a recent arXiv preprint, not yet peer-reviewed or widely cited. It is a well-constructed framework from a legitimate business-school source, but it is new work, not settled literature the way the machine-unlearning research we cite elsewhere is.

Why does it matter if nothing is actually being misused?

Because human relationships have always run on something the researchers call mutual forgetting — both sides lose detail over time, which the paper argues works as "a psychological safety net" that "supports dignity, second chances, and identity change."

A client who said something impulsive, got a fact wrong, or changed their mind six months ago benefits from the fact that your memory of that moment fades. A system that keeps it perfectly and forever removes that fading. At the extreme, the paper argues, this shades into narrative control: "the archive-holder acquires narrative authority" over what happened between the two parties, because one side has receipts and the other has recollection.

Does this change how a client experiences the relationship?

The paper argues it does, and not in the direction most product marketing assumes. People change their own behaviour — becoming more guarded, more self-censoring — simply from knowing a system will remember everything indefinitely, regardless of whether it is ever used against them.

The relationship the researchers theorise between memory power and trust is curvilinear, not a straight line. Some memory genuinely helps: the system remembers your preferences so you do not have to repeat yourself. Past a certain point, more persistent and more comprehensive memory starts to feel like surveillance rather than service.

What does the research say actually helps?

The paper is constructive rather than abolitionist — it proposes six design principles aimed at rebalancing the asymmetry rather than removing AI memory altogether. The four most relevant to a client-facing sales context:

> Memory transparency — making it visible to the person being remembered what is stored about them, and why a specific past detail is being surfaced now, rather than memory operating invisibly in the background.
> Forgetting-by-design — building in retention limits, decay, or an easy reset rather than keeping everything indefinitely because storage is cheap.
> Augmentation, not replacement — positioning the memory system as something that helps a person reconstruct and reflect on their own history, rather than a tool that quietly operates on them.
> Contextual containment — keeping memory scoped to the relationship and purpose it was collected for, rather than letting a casual comment about a personal situation get pulled into an unrelated later decision.

What does this mean for a tool that builds a knowledge graph of client relationships?

This is not an argument against persistent memory in a sales tool — the underlying capability is genuinely useful. It is an argument that the imbalance is real and worth designing for deliberately, instead of treating "we remember everything" as an unambiguous feature. Here is where Closer actually sits against those four principles, including where it falls short:

> Transparency, partially. The stored memory and the knowledge graph built from it are visible in the app, per client, to the account holder and their team. A client's memories can be exported as markdown. What does not exist is a client-facing view: the person being remembered cannot log in and inspect their own record.
> Forgetting, on request rather than on a timer. Deleting a client erases that client's entire memory scope and every row Closer owns for them, including the knowledge graph. There is an age-based retention sweep, but it ships off, and even switched on it does not purge the memory vector store. Closer does not forget anything on a schedule unless you configure it to.
> Containment, by construction. Memory is scoped per account and per client, and ownership is checked on every request before any client data is read. A detail from one client's conversation is not retrievable in another client's context.
> Augmentation is the design intent — prep briefs and Ask AI are for reconstructing what you and the client discussed. That still serves your side of the relationship first. The honest framing is that Closer narrows this asymmetry; it does not eliminate it.

If you want the imbalance to be smaller than the software makes it by default, the levers are behavioural as much as technical: tell people what you keep, offer them the export, and delete the record when the relationship ends.

Key takeaways

> Memory Power Asymmetry is a named 2025 research concept: one side of a relationship can reconstruct the shared past far better than the other.
> The researchers argue the imbalance is structural — it exists even when the memory is never misused and every privacy rule is followed.
> Mutual forgetting is doing real work in human relationships, and perfect machine memory removes it.
> The proposed fixes are transparency, deliberate forgetting, augmentation over replacement, and keeping memory scoped to its original context.
> Closer makes the stored record inspectable and exportable to the account holder, and deletes a client's whole memory scope on request — but there is no client-facing view, and nothing expires on a timer by default.