Part V
Why AI needs persistent client context
The devil in client knowledge lives in the details. You might tell a fresh ChatGPT session that the audience is value oriented. What you meant was that they are value oriented most of the time and behave differently on certain occasions. That qualification lives in your head because it has become obvious through months of work.
Experienced account teams and strategists accumulate hundreds of these details. They know what the client cares about, which fights matter, which words create trouble, which previous decisions remain valid and which apparently important documents everybody ignores.
Documents are not a curated memory
Uploading every client file sounds like the obvious solution. Unfortunately, client documents do not contain a clean hierarchy of truth.
A deck may put the central message on slide three and “ideas worth exploring” from an old workshop in an obscure appendix. The ideas were discussed, rejected and forgotten by everyone who attended. To an LLM reading the file, both pieces of text can look equally relevant.
The same collection may contain a current brand strategy, a superseded one, research with unresolved limitations, proposals that were never approved and a confident statement made by one stakeholder in a meeting. More context gives the system more material. It does not tell the system how much authority each item deserves.
Context needs status
Persistent client context should distinguish evidence, agreed decisions, working hypotheses, preferences and history. It should know when something was added, where it came from and whether later work replaced it.
Hypotheses should not silently become permanent client truth. Rejected ideas should remain rejected. Context-dependent statements need their conditions. A client preference can be important to the work without being an objective fact about the market.
This requires judgement from the people using the system. No automatic memory can infer every political or strategic nuance from documents alone. The software should make the status visible and easy to correct.
Memory belongs to the team
The knowledge cannot live only in the chat history of the person who discovered it. Agencies move people between accounts. Strategists are shared. Account directors leave. A project may pause for six months and return with a different team.
If the next person has to reconstruct the client's world from folders and private conversations, the agency pays for the same learning repeatedly. Important nuance is lost because the new team knows the headline and misses the qualification.
A shared memory should help someone understand what matters now and why. It should also show uncertainty, disagreement and change rather than presenting a single polished story.
Institutional memory can preserve institutional bias
Persistent context can make an agency smarter or simply more consistently wrong. Client teams accumulate beliefs such as “this client never approves humour,” “that audience does not care about sustainability,” “we tried that once” or “this stakeholder always says no.” Some are useful experience. Some are folklore created by one meeting and repeated until nobody remembers the source.
Memory needs provenance, confidence and opportunities for challenge. A belief should remain connected to the evidence or experience behind it. The system should make it possible to ask whether the condition still holds, rather than quietly presenting inherited prejudice as client truth.
Memory and context selection are different jobs
Persistent context does not mean using all context all the time. Good memory determines what can survive. Good context management determines what should be present for the task now.
An audience analysis, positioning decision and unresolved contradiction may matter to campaign strategy. A rejected event idea from last year probably does not. Pulling the entire client archive into every task recreates the document problem at a larger scale and gives irrelevant history another chance to distract the work.
The system should retrieve according to the decision being made while allowing a person to add the exception that matters in this case. Persistence makes knowledge available. Selection keeps it useful.
Forgetting can be useful
Good memory also includes a way to stop using something. Brands change direction. Evidence expires. People learn that an assumption was wrong. A system that remembers every statement forever will eventually become loyal to the past.
Perfect institutional memory seems like the wrong goal. Useful continuity preserves the decisions, evidence and nuance that help the next piece of work while keeping their status open to correction. That is closer to what experienced people do when they know a client well.