Build a company memory that ages and corrects itself

Joao Wenzel of Citadel Securities describes an institutional-memory system at AI Engineer.

Build a record from competing evidence

Strategy decks offer a top-down view; Slack, meeting notes and code changes can reveal a different picture.
A project can be green in a status deck while its comments expose an unresolved dependency.
Combine structured slide data with a vision model: each can recover details the other misses.
Preserve earlier monthly records and link renamed projects across time using people and goals as clues.

Update: record the reported status, supporting evidence, inferred status and rationale.
Age: date claims and reduce confidence as evidence gets older.
Observe: specialist agents inspect different materials and review one another’s findings.
Grade: compare a past snapshot’s warnings with subsequent project outcomes.
Correct: ask stakeholders for facts or reusable rules, and distinguish confirmed information from inference.
The model proposes changes; code and people decide how records are updated.

Wenzel says specialist workflows more than doubled the number of insightful project updates compared with a broad single pass.
In his retrospective test, green projects flagged by the agent were twice as likely to turn yellow or red. He presents this as a way to prioritize attention, not an exact diagnosis.
The talk does not give enough evaluation detail to establish how well the results generalize.

A missing signal is not proof that activity stopped: check whether the source itself was working.
Identifying the same client across systems without a shared identifier remains an open problem.
Agent-session traces are considered for engineers who opt in; sensitive workplace data still calls for careful access and privacy controls.
For high-stakes uncertainty, ask the right person at the right time. For low-stakes uncertainty, waiting for evidence may be appropriate.

“You start where the deck and the evidence disagree.”