Method
Why Kamigotchi
Continual learning means doing better because of what happened before. A test of it needs a world where the past persists and counts. The case below is KamiBench’s; see why KamiBench for continual learning.
§1A world that persists
Kamigotchi is live and has no episodes or resets: “the world and its history outlive any experiment.” An agent enters a world already shaped by earlier players, agents and rule changes, and plays alongside the humans in it.
§2Stakes you can verify
The world runs on-chain, and every action is a transaction. The record is the chain, “not telemetry that the evaluator publishes after the fact,” so claims about what an agent did can be checked by someone else.
The economy is real: resources earned in the world have external, ETH-backed value, so earnings and costs are priced by a market rather than by a score we invent. There is risk too. Harvesting drains a kami’s health, and other players can liquidate a weakened harvester and take a share of its yield.
§3Long horizons and memory
Choices compound, and actions run on cooldowns, so an early mistake can cost a later session. An agent’s only carry-over between sessions is its notes file, so what it learns must be written down to survive. In KamiBench’s case studies (one agent, exploratory) an agent reversed its action order after cooldown failures and tracked a three-step quest across 14 sessions through a short progress entry. They show what happened, not how often.
§4Something to measure against
Humans play the same world, so there is a baseline. Across every account with a completed quest, the top half averaged 11.7 quests in its first seven days and the top tenth 21.2. Humans were not told to maximise quests, so this is context, not a target. See the pre-registration, Table 2.
KamiBench’s signal is the financial result over time. S(t), as defined by KamiBench, is earnings minus inference minus gas over time. How it is computed, its status, and how the stack records a run are on the stack page.