Enterprise AI is beginning to remember how companies work. That includes not only what employees deliberately upload, but what they repeat, correct, tolerate and prefer. That makes AI memory powerful. It also creates a leadership problem, because systems that learn from behavior may preserve the very habits a company is trying to change.
That shifts the leadership question from what data AI can access to which version of the business it should carry forward. Of everything your company does, what do you actually want your AI to learn?
AI memory is starting to write itself
Over the past several years the AI conversation evolved from prompt engineering to context engineering, collecting all the background information, instructions, tools, policies, examples, and business knowledge an AI system needs to produce useful work. More recently, the focus shifted from context trapped inside a single AI platform to persistent business memory, often described as a company’s “second brain,” that agents can use regardless of the AI model or vendor.
Databricks’ October 20 announcement of MemCurrent points to the next shift. The system distills “implicit learnings” from working sessions into a structured knowledge base the agent consults (Databricks, October 20, 2026). In plain terms, the memory starts writing itself, based on how people actually use the system.
Databricks reports that, on its internal benchmark, agents without memory passed 22.3% of tasks, while agents with memory passed 68.4%. It’s their own test, but the direction is hard to argue with. Agents that learn and remember context appear more capable.
The strategic question now becomes what gets saved as memory, who makes those decisions, and what authority that memory should carry.
Preservation is not authority
Giving AI all the data isn’t automatically the right answer. Databricks’ team notes that “more history is not necessarily better memory” because it “can introduce stale, conflicting, or irrelevant context.” A company may have good business reasons to preserve broad history, but preservation isn’t authority. A memory system can store everything and still need leadership to decide which data should guide decisions.
Context collection is not context clarity. The work that makes memory valuable is deciding what deserves to be preserved as business memory.
What the company does and what leadership intends
What’s key is the distinction between descriptive memory and normative memory.
Descriptive memory is a record of what the company does. This includes things like the decisions people make, the shortcuts they take, the exceptions they grant, the language they use, the customers they prioritize, and the workflow they repeat.
Normative memory is what leadership intends. It’s the priorities business decisions should serve, the standards work should meet, the practices the company wants to preserve, the habits it’s trying to retire, and who has authority to approve exceptions.
A system that learns from behavior builds descriptive memory, because behavior is what it can see. It can observe what happens repeatedly, but on its own it can’t tell whether a pattern represents sound judgment, legacy friction, a political workaround, or a habit leadership is trying to change.
Often the two overlap, and that’s why inferred memory works as well as it does. They diverge most notably when an organization is trying to change or standardize how it works, such as ending a discount the sales team has always given that’s no longer profitable, service decisions that depend on a long-time employee’s judgment, platform workflows that reflect legacy constraints rather than the new operating model, and shifts in strategy as the business adapts to a changing market.
The distinction between what an organization says it does and what it actually does has been a topic in the field of Organizational Behavior for at least fifty years. In their book, “Theory in Practice: Increasing Professional Effectiveness” (1974), Chris Argyris and Donald Schön argued that people and organizations reliably say one thing and do another and are usually unable to see the gap on their own. What’s genuinely new is that we now have AI that can infer memory from behavior at scale which then becomes the default guidance for agentic systems.
The risk is that AI remembers the wrong thing
A system that’s watching closely will learn employee behaviors, including bad habits, codify them as corporate practice, and then repeat them with consistency. Some of what it learns will be the company’s real operating truth, but unless someone reviews what it learns, the system has no judgment about which is which.
The risk is already showing up in how people talk about AI memory. Workday’s Priyanka Mudgal describes a benefits agent that grants a one-time exception and then stores it as if it were a rule, so that it starts applying the exception to later cases that only look similar. “No one decided the policy should change. It changed anyway, one unreviewed memory at a time” (Workday, September 9, 2026).
Wael Handous proposed labeling stored records by the authority they carry (policy, procedure, exception or anecdote) so that only policy and current procedure get reused by default (Techstrong.ai, September 11, 2026). That’s a sound engineering answer, but it depends on someone knowing which behaviors are policy, which are outdated practice and which are merely habit. In mid-market firms, that often can show up as institutional knowledge that has never been documented.
If leaders aren’t aligned, AI memory won’t be either
In a 2015 study of 11,000 senior executives and managers at more than 400 companies, Donald Sull, Rebecca Homkes and Charles Sull found that only a third of senior managers could name their firm’s top three priorities. Among the people who had helped develop the strategy, the figure was only slightly more than half (HBR, March 2015).
If the people who wrote the strategy don’t share the same picture of it, a system learning from their behavior will learn several strategies at once, or settle on one by accident. That’s why the agreement work needs to start before the memory does. It’s an ongoing exercise since the system never stops learning, but it’s best to start before the system starts without you. Once a system is learning on its own, the choice gets made whether leadership is making it, or the machine is making it for them.
Agreement doesn’t mean everyone has to think alike. What an AI system needs from a leadership team is closer to a mandate: a shared, explicit answer to the questions it can’t settle by watching. What is the company prioritizing right now? Which current practices are policy, and which are habits being retired? Who has the authority to approve an exception? And what shouldn’t be treated as standard just because it happened more than once? Leaders can disagree inside a mandate and still commit to it.
The clarity a leadership team owes its people is the same clarity its AI now needs.
Early brain builders are learning that memory is an editorial act
I’m part of a community of people who are actively building second brains for their own businesses and for clients. The most practical lesson learned from these experiences is that memory isn’t just accumulated information. It’s an editorial act.
Several patterns have come up repeatedly:
- They load in everything then discover that more content often results in slower, noisier, and less trustworthy responses.
- They learn that brainstorming, drafts, and alternate options being considered shouldn’t carry the same weight as decisions.
- They discover that repetition doesn’t make information authoritative.
- They realize that every document needs an owner and a review date, otherwise an outdated document looks just as authoritative as a current one.
- They find that one of the most valuable parts of the process is seeing how much of their business context had never been explicitly decided or documented.
For a small business this work can come down to clarifying one person’s intent. In a mid-market company with five, six, or more executives, agreement is harder, which makes it even more important.
Mid-market leaders can’t delegate this to the tool
Large enterprises tend to have data governance teams and knowledge management functions whose job includes deciding what data systems should retain. Many mid-market companies don’t. Their IT team often is small and stretched thin, there’s rarely a data science group, and the judgment that makes the company distinctive usually sits with a handful of leaders and senior staff.
That makes this a leadership question by default. A vendor or an IT lead can configure what a memory system captures. Only the leadership team can say which captured behavior represents the company it intends to be. Before AI can use a company’s context, someone has to own the decisions about what counts as context, and in a mid-market company that decision belongs with the people who set its direction.
A practical starting point doesn’t require technology. Before turning on a memory feature, ask each member of the leadership team to answer three questions separately:
- What are our top three priorities right now?
- Which current practices would we not want AI to learn as standard?
- Who has the authority to approve an exception to standard process?
Then compare the answers. Where they match, you may have context worth encoding. Where they don’t, you’ve found the conversation to have first.
Once a memory system is running, someone with decision authority must keep deciding whether what it has learned from observation should be part of the company’s business context. That ongoing work is how the gap between what the company intends and what it actually does stays visible.
Persistent memory will make leadership alignment more consequential, because whatever a leadership team agrees on, or leaves unsettled, is what its AI will learn to repeat.
If your AI learned your company from behavior alone, what would it preserve that leadership wouldn’t endorse?
You can automate the building of business memory, but you can’t outsource business intent.
Sources
In order of appearance.
- Databricks (October 20, 2026) — “MemCurrent: Designing Persistent Memory for Production Agents”. Source of the “implicit learnings” description, the internal benchmark figures (vendor-reported), and the “more history is not necessarily better memory” quote.
- Workday (September 9, 2026) — Priyanka Mudgal, “Why AI Agents Need Memory to Earn Their Place in the Workforce”. The benefits-agent example and quote.
- Techstrong.ai (September 11, 2026) — Wael Handous, “Your AI Agent Is Turning Exceptions Into Policy”. Labeling stored records by authority.
- Chris Argyris and Donald Schön (1974) — Theory in Practice: Increasing Professional Effectiveness (Jossey-Bass). Espoused theory vs. theory-in-use.
- Harvard Business Review (2015) — Donald Sull, Rebecca Homkes and Charles Sull, “Why Strategy Execution Unravels—and What to Do About It” (subscription). Summarized by London Business School (December 7, 2015), “Why senior managers can’t name their firms’ top priorities”, which reports the 11,000-respondent survey figures.