Swarm Lab · Experiment 04 · Voyager 失忆岛
The island that forgot the manual
Does experience travel—or merely repeat itself?
Three learners arrive on the same islands. One packs nothing. One packs working skills. One also keeps a record of what failed. On the next island, some familiar rules are wrong.
Your challenge
Find out when experience transfers—and when it becomes stale advice.
- Play through all 120 attempts. Islands change after attempts 40 and 80.
- Set Changed rules to 0. Existing skills should now need no extra failure memory.
- Set it to 4 and compare repeated mistakes. All learners still get 120 test attempts.
First time here? The buttons explained
Play starts or pauses. Step advances one frame; the time slider rewinds. Changing a control prepares a new run; press Play again. Reset replays the current settings. Seed chooses a random world, not difficulty; keep it while changing one rule, then try Next seed. Share saves your setup in a link. The downloads let you inspect the data.
What every parameter means
- Seed
- A repeatable random world, from 0 to 4294967295. The compared arms use the same seed.
- Changed rules per island
- The number of the four context rules that change on each new island. 0 means a familiar world; 4 means every rule changes.
Make a prediction. Then press Play.
Model 1.1.0 · simulation onlyEach learner gets exactly 120 action attempts: forty on each island. A dot is one attempt, green for success and coral for failure. Dashed lines mark the island changes. The dashed chart line is the first arm. Controls change both arms except the named treatment; the seed is shared.
Read the result: The chart is cumulative success rate. Skills arms start with prior knowledge; the fresh arm does not. The clean extra-memory comparison is Skills versus Skills + failure memory.
Loading local simulation…
Now break the story
Set changed rules to zero. How much of the apparent intelligence was just useful luggage?
Move one control, keep the seed, and compare again. The batch below is a fixed baseline—not a live summary of your current settings.
Observed in this model · 32 matched seeds · all runs retained
A result, not just an animation.
Skills help. Remembering failed actions helps further while rules stay fixed within an island. The advantage is a designed exclusion rule, not evidence of reflective consciousness.
| Arm | Mean | Seed range |
|---|---|---|
| Fresh learner | 72.2% | 50.8% – 85.0% |
| Skills | 87.7% | 74.2% – 95.0% |
| Skills + failure memory | 93.5% | 90.8% – 95.8% |
Ranges describe these runs, not confidence intervals for real-world systems.
Under the hood · the actual rules
Four contexts each require one of four actions. The first two changed contexts (default) rotate their correct action on each new island. Fresh learners clear successful-action memory at each crossing. Skill learners retain it. Both explore uniformly after a failure. The third learner additionally excludes actions that failed in the current island/context. All arms see identical task order and draw keys. Skills start with four policies learned in a previous world; that training cost is outside this comparison.
Read this model’s source · Read invariant tests · Reproduce the receipts
Where the analogy stops
No LLM, Minecraft, live agent or Voyager software runs here. These are explicit rule-based learners. Failure memory has an advantage by construction when rules stay fixed within an island; this model tests that assumption, not general intelligence. Equal test budgets do not imply equal lifetime training cost.
Sources, methods & reproduction
Inspired by Robin’s diary. Background: Voyager · original research project. This is an original teaching model, not a replication of that research.
Experiment notes · Run locally with Node · Version & source hashes