Agents, Personal Guidance, and Memory: How We Think About Building FinCatch

A deeper question

When people talk about the future of AI, they often focus on the model itself, which benchmark it won, how fast it is, or how natural it sounds. But for a financial system, the deeper question is not just how to generate a good answer. It is how to build something that can act like an agent over time, offer useful personal guidance, and remember enough to understand how an investor's thinking evolves.

That has increasingly become the way we think about FinCatch. Over the last year, we have been building the system piece by piece around a simple idea. A serious financial assistant cannot live only in the visible chat. It needs a deeper structure underneath. In early April, we made a major redesign to our product so that several threads already present in the system, our financial world model, our agent workflows, and a new personal memory layer, could come together as one more unified architecture. That architecture now powers an end to end agentic equity research pipeline for stocks.

The three layer mind

The way this architecture now makes sense to us is as a kind of three layer mind. At the base is the market layer, a financial world model built as an event centric knowledge graph that captures what is happening across companies, sectors, guidance, analyst revisions, and the relationships that connect those events into narratives over time. You can see that foundation in From Facts to Foresight and A Collective Unconscious for Markets. Above that sits an unconscious personal layer, a separate memory graph being built to capture what each user has been exploring, asking, and gradually focusing on as they work through the market, as outlined in How FinCatch Remembers What You Think About the Market. At the surface is the conscious layer, the active chat, the current tasks, retrieved documents, and the short term context that lets the system respond right now. Under the hood, that means two graphs, one for the market's world and one for your world, with the conscious layer sitting on top and stitching them together in each interaction.

This framing matters because in practice, financial work happens across all three layers at once. Documents, transcripts, market data, and chat history are still useful, and techniques like retrieval and search remain important. But alongside those, there is a deeper need to model what the market is doing structurally and what the user is thinking cumulatively. That is where the world model and personal memory begin to matter.

What recent AI signals confirmed

That is one reason the latest moves from larger AI labs felt familiar. According to Google I/O 2026, Google made it clear that the center of gravity is moving from one shot chat toward systems that persist, monitor, and act across time. That is not the same product problem FinCatch is solving, but the conceptual direction is similar. AI is increasingly being framed not just as something that answers questions, but as something that works in the background, keeps track of changing context, and helps users move through a stream of decisions rather than a single prompt.

Something similar shows up in Anthropic's research on how people ask Claude for personal guidance. In a study of one million conversations, Anthropic found that about 6% of Claude conversations were people seeking personal guidance, and within that group personal finance accounted for 11%, which is a useful signal that users increasingly want AI not just to inform them, but to help them think through money decisions. It also highlights something especially important in finance. A guidance system has to do more than sound helpful. It has to help users think more clearly, especially when the right answer is not simply to reinforce the first view on the table.

What this means for FinCatch

For FinCatch, those signals are less about imitation and more about confirmation. People do not only want an AI that can summarize the market. They want one that can help them decide what to pay attention to, what to question, and what to do next. In finance, that means the system has to reconcile three things at once, what the market is objectively doing, what the user has previously been exploring, and what is being asked right now.

That reconciliation is where the three layers become more than a metaphor. When a new earnings release, guidance change, or analyst revision enters the market graph, the system can compare it against the user's personal memory graph and see whether the new evidence supports, weakens, or complicates a direction the user has been pursuing. The conscious layer then decides how to surface that tension, as a prompt, a suggested line of inquiry, a scenario to test, or a task to run. Instead of merely retrieving context, the system can begin to reason about the relationship between market reality and investor focus.

How this looks in practice

FinCatch today focuses on stocks, where the layers already come together in concrete ways.

When you ask about a company around earnings, the research agent does not just fetch the latest transcript. It walks the event centric market graph across that company's supply chain, peer set, and sector relationships to build a set of plausible paths before the print. According to our work on scenario thinking, the agent maintains upside and downside scenarios with explicit triggers, so when new information arrives it knows which branch of the tree you are now on rather than treating it as a fresh question.

Under the hood, this flow runs on the same agentic equity research pipeline we described recently, where specialised agents hand work off to one another rather than a single prompt trying to do everything at once.

When management reports, our guidance extraction system pulls forward looking statements from transcripts and press releases, normalises them into a structured time series, and grades whether revised guidance is stronger or weaker than before. That means the system can show exactly how a new guidance print compares to what management said the prior quarter, and how that changes the assumptions driving your analysis rather than just summarising what was said.

From there, the agentic financial modelling framework can build or update a model around the same stock, with projections tied directly back to the events and guidance signals that moved the story. Each model run is traceable to specific market events, so you can see how assumptions changed from one version to the next rather than working from a static template.

All of this runs on top of an infrastructure designed to keep agent state alive across requests, not just within a single session, as described in Claw on Cloud. That is what makes it possible for the system to pick up where you left off across days of research on the same stock, rather than starting from scratch every time.

How the pieces fit together

This is also why we have spent so much time on pieces that may look disconnected from the outside but are actually parts of the same design. Our work on harness engineering was about making agent behavior reliable in the real world. Our work on scenarios and agentic financial modelling was about helping the system explore what could happen next instead of stopping at static analysis. Our work on management guidance extraction and the event centric knowledge graph was about representing the market as a structured, evolving system rather than a pile of documents. All of these pieces now live inside one agentic equity research pipeline, so that asking a question about a stock can trigger the right sequence of agents, tools, and checks without you having to orchestrate them manually.

Traditional recommendation systems, like those behind early video platforms, mainly use two approaches: collaborative filtering, which learns from user behavior patterns through user based filtering, where people similar to you also watched Y, and item based filtering, where people who watched X also watched Y, and content based filtering, which recommends items with similar attributes like genre, director, or keywords. FinCatch's unconscious personal memory is different because it is being built to model how your own focus, questions, and direction of inquiry evolve over time across individual stocks, not just what similar users clicked on before. According to the design outlined in How FinCatch Remembers What You Think About the Market, this means the system maintains a live view of where your attention sits on each name, with a versioned timeline of how that focus has shifted, rather than a flat list of what you have opened.

Seen this way, the personal memory layer is not an optional personalization feature added on top. It is the missing middle between the market's collective structure and the chat interface where you experience the product. Without it, an agent may know the world but not the investor. With it, the system can start to notice patterns in what you care about, how you frame questions, where you tend to look first, and when new information should trigger a genuine rethink rather than a routine update.

What good guidance requires

That shift also changes what guidance means. Good guidance is not about giving the most comforting answer, and it is not about pretending the system knows with certainty what someone should do. It is about helping you see something you might otherwise miss, a shift in guidance that changes a key assumption, a pattern across peers that puts a single name in a different light, or a fresh event that should change the weight placed on an old line of analysis. That is part of what made the Anthropic findings on how people ask Claude for personal guidance so interesting to us. They point to a broader shift in what people now expect from AI, especially in areas that involve judgment and real stakes.

What comes next

This is the direction FinCatch is now being built around. The next product step is bringing the personal memory layer fully online, so the system can remember not just what happened in the market, but also what each investor has been trying to understand about it across sessions and across stocks. As described in How FinCatch Remembers What You Think About the Market, that means a structured memory graph with a live view and a versioned timeline for each stock you care about, connected back to the same event centric world model the research agent already uses.

The broader AI industry is moving toward agents and guidance. The lesson for finance is that neither works especially well without memory, and that memory becomes much more powerful when it is structured, layered, and continuously reconciled with the world. That is what we are building at FinCatch, and why we think the approach is worth getting right.