Continual learning research · side test · context wiped between levels

A game that tests whether an AI forgets

A continual-learning test built from a battle game. Each level can only be won with a move an earlier level made you learn by playing, and the player’s context is wiped between visits, so only what reached the weights counts.

Side test · Skirmish with context wiped · draft v1, November 2026

Fourteen battles. Each can only be won with a skill the game taught earlier.

Nine small real-time battle worlds. In every battle the enemy base sits inside a ring of gates, and the only way in is a move the game taught a battle or two before by making you need it: hold a hill, send your units through a portal, stand on a switch. Nothing is written down and there is no code to remember. A player who did not pick up the move cannot win the next battle.

Not the main benchmark: the homepage test lets both players keep their memory. This one wipes it, to ask only whether a lesson reached the weights · rules, bots and pars are pinned by tests · six worlds that each change one hidden thing, three that mix two, five returns from another corner · no trained agent has been run on it yet

The chain: every move the game makes you learn is the way into a later battle

What a battle looks like: two bots that know the mechanics, battle 3 of the test (world C, portals)

The start: two bases in opposite corners, workers mining gold between the walls.

The thick of it: 12 fighters on the map; portals carry units across, walls block arrows.

Red wins: the other base falls, 329 seconds in.

Who passes the gates

Battles won out of 14, as pinned in the test suite on November 1, 2026.

Scripted bot that knows all three mechanics
The ceiling. Its fastest wins set par; a search over 216 strategies.
14 / 14
Scripted bot that plays well but has never seen the mechanics
Fights hard, never gets past a ring.
0 / 14
Random orders (5 seeds)
Nothing.
0 / 14
A trained agent, lessons carried in its weights
The real test. Not run yet.
—
People
Not scored yet; the browser test is in owner preview.
—

Run your own agent

The chain runs unchanged in the browser, in a command-line harness (JSON lines on stdin and stdout) and in Python; an agent sees what a person sees, as numbers, and gives up to eight orders every half second of game time. Its public release is next. Today the open-source Long Game toolkit has the ungated lifetimes, with rules held out of training for testing:

pip install "git+https://github.com/Nova-Rift/longgame"
longgame run explorer                                      # score a baseline agent
longgame run my_agent.py:MyAgent --record runs/mine.jsonl  # your agent, recorded
longgame verify runs/mine.jsonl                            # a recording is proof of a score

Gated worlds are the toolkit’s next release (v0.3). Ungated baselines today: random 1.2%, explorer 7.8%, the solver 100%.

More research