Transactive memory

Transactive memory experiment

A transactive-memory directory estimates which member knows about which topic. The earlier hive experiment folded that estimate from grounded deposits and citations, then used BidReason::Knows to invite a member who had not yet taken a position. !defer let a member decline a topic and change that estimate. This page records the idea and its measured outcome. The pure directory fold remains available for task division and other callers. Its episode attention wiring, Knows bid, and deferral were retired.

Why it was tried

A roster says who is present. It says little about who has the useful fact for a particular question. The proposed directory tried to learn that distinction from the transcript without adding storage or trusting a role label as proof. A citation was evidence that another member's contribution mattered; a bare statement was weaker. The fold could be reproduced from the shared log.

The risk was circularity: a frequent speaker could look like an expert because their messages were available to cite. The experiment therefore measured the correlation between directory weight and speech share alongside accuracy.

What the runs found

In the recorded simulated rooms, adding the directory did not improve accuracy beyond the base policy's interval. An evidence-first opening made the Knows bid reachable and often selected it, but accuracy was 65.8% against 66.3% for the same policy without the directory. Deferral also had no clear gain. In the live matrix, no turn was awarded through Knows, and agents did not use !defer despite seeing it in their move list.

Those results do not show that transactive memory is useless for every group. They show that this implementation had no demonstrated benefit for the tested rooms and made the protocol larger. ADR 0027 records the removal of its episode wiring. The full data and limitations remain in the expert delegation experiment and live matrix.