Your AI Agent Knows Everything—and Understands Nothing

“We should join my parents on their trip to Ireland” — this one seemingly innocuous statement sent shivers down my spine.

My wife and I travel extensively. We know what we like. My wife’s parents, on the other hand, rarely travel further than a few hundred miles from their home and have left the country together a grand total of once — for our wedding.

To top it all off, this trip was a Christmas gift from my father-in-law to my mother-in-law so she could go and visit her family, perhaps one last time.

I could see this trip unfolding in a single word: disaster. How in the world would we synthesize wildly different experiences and expectations so that we could have an amazing trip — or at least not hate each other at the end of it?

Like any self-respecting techie, I reached for technology — specifically for AI.

But what I didn’t expect is that my little experiment in vibe-coding an AI-powered family trip planning app would teach me almost everything I need to know about applying AI within enterprise IT.

The More You Feed AI, the Dumber It Gets

Most enterprise AI deployments follow a predictable pattern. Organizations start by giving an agent a set of instructions and connecting it to an information source, whether that’s a RAG (Retrieval-Augmented Generation) framework, an existing knowledge base, or even an MCP server. Next, layer in an LLM and let it do its thing.

The problem there is that LLMs at their core are dumb. They don’t know how to prioritize all the information they have at their disposal, so they tend to treat every piece of context equally. A human has to add a layer of curation, teaching the model what’s important and what’s not. Without curation, you get AI that knows everything and understands nothing.

The Three Types of Memory That Matter

Effective enterprise AI curation means making the most of three specific types of memory.

The first is institutional memory, which can seem pretty basic at the beginning. When someone says “financial services,” the agent knows they mean the company’s Financial Services division and not the whole industry. That becomes persistent organizational knowledge filled with definitions, preferences and conventions that don’t often change. As this extends into institutional knowledge around strategic priorities, key initiatives, and organizational dynamics, it becomes a rich source of institutional context.

Next is action history, which focuses on significant decisions, tasks and events. When a service ticket is filed or a system deployed, the agent recognizes that particular action and records it into the action history. This becomes the historical record that stitches together organizational context.

Finally, there’s short-term conversational context. Think of it as the moment-to-moment interaction with an agent. It’s useful in the moment, but tends to lose relevance quickly.

Taken together, these three types of memory create the weighting system that generic AI models are missing. Now when someone tells an agent about the business, they’re classifying and prioritizing all that memory and curating the information that’s important. This forms the core of what AI should deliver: not just domain data, but domain judgment.

What Curated Memory Looks Like at Scale

But enough with the framework, what does this look like in practice? Here’s what we’ve discovered in building these agents ourselves.

One common IT scenario is sending a trouble ticket to a help desk agent. Say your Outlook isn’t working, so you type in a description of the problem and wait for the agent to review and suggest a fix.

But with curated memory working in your favor, a better process could involve taking a screenshot that shows the Outlook error and uploading that to the agent. Now the agent (1) draws on institutional memory to understand your working environment; (2) checks the action history for related incidents; and (3) applies contextual judgment for a specific solution, not just a generic answer.

The result is an intelligent agent that doesn’t have to guess the answer based on a screenshot. It’s now actually interrogating, looking at all the information currently running and delivering a more useful response. The agent could even expand into a network or swarm effect, looking at other users in the system to see if the Outlook problem is only you or an enterprise-wide problem.

The contextualization of the history or memory is the difference-maker. If you don’t curate your memory effectively, you’ll fall behind those who do. It’s essential to have an architecture that knows how to manage that data over time and understand what to keep, what to surface, and what to let go.

Back to the Trip

So, how did my AI-powered trip planner change my view of AI in enterprise IT?

What I built was an app that acted as our personal trip concierge and began by “interviewing” each participant. We all explained what mattered to us on the trip: what was a must do and what we could skip. More importantly, it asked us about our “why” — why was something important to us, what did it mean to us.

Using this information, it did two things. First, it curated a trip plan that was balanced to deliver something for everyone — we could all see our desires and preferences represented in the plan it produced.

But, of course, that first itinerary was just a draft. There were still many questions to answer.

And that’s when the real magic happened. We asked the agent about a hotel or an attraction or a drive, and the answers it gave us were enriched with the context of our unique situation: “It would be a long drive for the kids, but my father-in-law would love the castle (and the unique coffee house next door) — and this could be just the spot for my wife to get that massage.”

Full of this rich understanding of what was important to us, it was able to help us plan and refine our trip in a way that I don’t think would have been possible in any other way.

And it was in one of those first moments that I understood what we needed to build for our enterprise customers: intelligent systems that were so laden in organizational, transactional, and personal context that every answer and every interaction would be like a fingerprint: completely unique to that moment and interaction that it would deliver a type of value that could simply not happen any other way.