How Citadel Securities builds a memory that catches troubled projects

Most companies have someone who remembers the reorganizations, relationships and politics behind formal decisions. Wenzel calls that person a kind of company world model: useful, but difficult to scale and rarely written down.

At Citadel Securities, Wenzel says agents already help employees with work across the organization. Leadership then asked whether they could also help define and refine company strategy, which requires understanding the company’s current position, history and likely next steps.

The institutional memory must capture relationships and project histories, including why teams stopped working together or a project was rescoped. Wenzel contrasts formal decks and strategy reports with noisier meeting notes, messages, code changes and agent-session traces.

Formal reports can contain what he calls watermelon projects: green on the outside but red inside. Operational sources carry useful signals too, though they are noisy. The company model is where evidence from both directions is brought together.

Wenzel borrows the idea of a state estimator from robotics. A belief becomes less certain as supporting evidence ages, and new observations update it. He says the system should consider how recent a signal is, whether its source is trustworthy and whether other sources agree.

Extracting even a status slide is difficult. Parsing PowerPoint XML may expose data but makes visual concepts such as color awkward; a vision model may struggle with a screenshot copied into the presentation. Combining the structured file with the vision model’s reading gives a fuller picture.

Projects also change names, so records must be linked over time. A five-day-old deck can disagree with a later Slack thread or a dashboard image; a longitudinal audit trail makes those changes and contradictions traceable.

His proposed loop is update, age, observe, grade and correct. Grading compares observations with expectations, and correction brings stakeholder feedback back into the system. Agents participate throughout, but code and people decide how records are updated.

A retrieval system can find a few messages and slides and answer that a project is green while missing important context. Wenzel’s approach constructs an ongoing record instead of relying only on whatever documents a query retrieves.

Each month the system combines a prior record with the new status deck. The team has about three years of records. It can infer that a renamed initiative is the same project by comparing the people involved and the continuity of its goals.

An agent also separates a reported status from its evidence. A deck may label a project green, while a comment about an unresolved dependency supports a yellow flag. The update records what was reported, the evidence, the inferred status and the reason.

The system does not rewrite previous months. Corrections feed forward instead, preserving historical snapshots for evaluation. That history also shows how quickly different facts become stale.

Delivery dates are one example: a launch scheduled for Q2 may move to Q3 and then October. Across three years of records, Wenzel says half the dates changed within two months, and 90% of those changes moved the date later.

Associating a date with each claim lets an agent take recency into account when answering questions about a project, a person or the company’s intentions.

Initially the team tried asking one model to examine the whole corpus at once. Wenzel says it lost detail. The team instead divided materials among specialist agents, added agent review and measured what came back; in the example, insightful project updates more than doubled.

The next test asks whether the model can identify trouble before a project’s formal status changes. The team reconstructs a past snapshot, runs its analysis using information available then and compares its warnings with subsequent labels.

In that backtest, projects marked green but flagged by the agent were about twice as likely to turn yellow or red. Wenzel says this does not predict exactly what will go wrong; it can help people choose where to look.

Human feedback is another part of the loop. Wenzel recommends meeting stakeholders in the tools they already use, such as Slack or email, and making it easy to respond. A correction that expresses a rule can improve future outputs as well as the current one.

The record can distinguish something inferred by an agent from something a person has confirmed. That distinction preserves the origin of a claim when the system is used later.

Bottom-up evidence includes meetings, Slack, code changes and agent sessions. For a Slack thread, the system can extract the claim, date, participants, related project and link to evidence; pull requests show what changed.

For engineers who opt in, agent-session traces can add the history behind a pull request: how the request evolved, failed tests and decisions made along the way. Wenzel says combining those traces with other sources gives a fuller picture.

Signals can fail. A speedometer going blank does not mean a car has stopped; similarly, a broken data source can look like a quiet project. The team learned to track when it last attempted a fetch and when the source last worked.

The resulting record can say more than green: a launch date slipped twice, the evidence suggests yellow and people are discussing failing tests. Leaders can start conversations where the deck and the evidence disagree, and managers receive a weekly view of what advanced or went quiet.

Wenzel says the same context-layer pattern could be extended to client information, combining internal and external sources to understand client interests and anticipate needs. One unresolved problem is matching records for the same client across systems without a shared identifier.

The right human intervention depends on confidence and stakes. For low-stakes changes the system can apply confident updates or wait for more evidence if uncertain; high-stakes cases require care over whom to ask, when and how.

Wenzel closes with the five-step loop: update, age, observe, grade and correct. He recommends evaluating against what happens next and designing how agents and stakeholders work together, instead of assuming that indexing every document captures the context between the lines.