THE LABORATORY / PREDICT · INTERVENE · EXPLAIN
Learn by changing the system.
Three experiments in memory, feedback, and learning from the wrong success signal. Make a prediction, change one thing, and inspect the evidence.
Start with storage on and retrieval off. Does writing a memory make it available at the door?
No stored observations.
Answer not attempted yet
Answer not attempted yet
Inspect the rules & observation history
Each event takes one slot. New events evict the oldest. Storage writes matching events to a separate archive. At step 10, retrieval may insert one east-door record, evicting a slot if needed. The answer uses the highest visible version in active context. The lock checks the actual current code. Every read and write has equal unit cost here; these are not tokens or prices.
TRANSFER THE INSIGHT
What must a memory system get right?
Separate what gets written, what gets retrieved, and which version is trusted. A larger archive can contain more wrong answers. What evidence would you need to choose a memory policy for a real task?
Compare with a worked explanation
With three slots and a rotated code, storing codes and retrieving the first match returns 4821: wrong. Retrieving the latest east-door version returns 9063 with three writes and one read. Storing all nine observations returns the same answer at more write cost. With nine context slots, even the control can see the latest version. The benefit depends on the context budget and the retrieval rule.