Do Lessons Travel?
Can verified notes, components, and decisions from an earlier project improve outcomes on a genuinely new task?
RESEARCH AGENDA / PROPOSED STUDIES
Can an Organization Actually Compound What It Learns?
The white paper is a vision. The research program asks what evidence would be needed to support, refine, or reject its central claims.
Which Tasks Should AI Own?
Compare human, single-agent, and coordinated-agent workflows across complexity, cost, and error.
What Counts as Progress?
Do failed experiments become useful knowledge, or do they simply consume resources?
A FIRST CONTROLLED EXPERIMENT
One Task Family. Four Conditions.
Match tasks and budgets, verify the actual outputs, and preserve failures so that improvements cannot be inferred from agent confidence alone.
COMPARISON CONDITIONS
Baseline vs Shared Knowledge
One agent without memory; one with curated project memory; multiple agents sharing knowledge; and a condition with independent human review.
MEASURES
Success, Cost, Time, and Transfer.
Use independent tests and a task set not seen while creating the memory. Record retries, human work, wrong transfers, and negative findings.
RESEARCH DISCIPLINE
What Would Change Our Minds?
If shared knowledge makes new tasks slower, less reliable, harder to maintain, or more expensive, that result belongs in the paper too.
Preserve the Evidence.
Source revision, task definition, model choice, validation criteria, errors, and human intervention.
Challenge the Transfer.
Don't let agents memorize answers or reuse information that doesn't fit the new task.
Publish What Happened.
Include limitations, incomplete attempts, failures, and replication details where possible.
Research directions are proposed, not presented as completed experiments. See the complementary coordination protocol on Scholar.