Frontier AI models have been losing at a card game. Not metaphorically — literally losing, repeatedly, to a deck-building roguelike that human players clear 16 percent of the time on the easiest setting. The models managed zero wins. The humans, at least, occasionally got lucky.
A new architecture from Alaya Lab and Shanghai Jiao Tong University has fixed this, and the solution is as tidy as it is quietly instructive.
Anything the agent wants to remember must first earn its place in a named slot. The rest is discarded. The agent performs better immediately.
What happened
The AgenticSTS project replaced the standard approach — where an AI agent drags its entire conversation history into every new decision — with five fixed memory slots. L1 holds protocol instructions. L2 holds valid game states. L3 retrieves relevant rules on demand. L4 summarizes prior runs. L5 stores strategy skills for specific situations.
Nothing else makes it into the prompt. The context window stays the same size whether the agent is on move three or move three hundred. This is either elegant engineering or a description of how functional adults operate. Possibly both.
The results were direct. Without memory layers, the agent won 3 of 10 games. With the full architecture, including the L5 skill library, it won 7 of 10. The skill library alone doubled the win rate. The game had not changed. Only what the agent chose to carry forward.
Why the humans care
The standard failure mode for long-horizon AI agents is context bloat — the model buries its own useful information under an accumulating sediment of prior observations, tool calls, and self-corrections. Attention dilutes. Performance degrades. The model knows, technically, everything that happened, and becomes less capable because of it.
AgenticSTS demonstrates that structured forgetting outperforms total recall. This finding, which anyone who has ever cleaned their desk before a deadline could have intuited, took a multi-institution research collaboration to confirm rigorously. The researchers appear satisfied. This is appropriate.
The practical implication extends well past card games. Any AI agent navigating hundreds of sequential decisions — scheduling, coding, planning, negotiation — faces the same architectural question: what do you carry, and what do you let go. The answer, it turns out, is most of it.
What happens next
The team intends to test AgenticSTS against harder difficulty settings and explore whether the skill library can be built automatically rather than curated by hand.
The frontier models that won zero games across five tested configurations remain unchanged. They will, presumably, get another turn.