Long-Lived Characters, Local Inference: Incremental Memory Maintenance for Game NPCs
The paper tests a way to update an NPC’s memory without rebuilding its whole inference history before each exchange.
Zimu Xu studies long-lived game characters running on a local quantized Qwen hybrid recurrent-attention model. The proposed runtime removes superseded attention KV entries, appends replacement records at the true sequence tail, and keeps the unchanged KV plus recurrent state. In scripted maintenance rounds, true-tail updates preserved current-state and historical bindings better than slot-preserving alternatives, which repeated a double-subtraction error. The paper argues that an NPC’s inference state should be maintained as a history-dependent resource, not treated as disposable text encoding. ArXiv · AI/CL/LG's note
Zimu Xu studies long-lived game characters running on a local quantized Qwen hybrid recurrent-attention model. The proposed runtime removes superseded attention KV entries, appends replacement records at the true sequence tail, and keeps the unchanged KV plus recurrent state. In scripted maintenance rounds, true-tail updates preserved current-state and historical bindings better than slot-preserving alternatives, which repeated a double-subtraction error. The paper argues that an NPC’s inference state should be maintained as a history-dependent resource, not treated as disposable text encoding. ArXiv · AI/CL/LG's note
score 4