The current line is Void Realm — focused on testing a single character's mind, memory, and personality.
All pre-2.0 versions — the functional-testing era.
Core upgrade: bi-temporal knowledge-graph memory + modern retrieval/forgetting + goal-driven colloquial dialogue. Characters finally remember, stay consistent, and stop being goldfish.
Read entry →Post-playtest fixes on the Void Realm stage: a smooth follow camera, object occupancy so lying down actually lands on the bed, cleaner action text, and — the one cognitive change — emotion valence that finally moves instead of staying frozen at zero.
Read entry →Builds a single-agent isolation world (the Void Realm) so one mind's memory, emotion, and self-image can be observed cleanly, and adds Deep Replay — a way to look step-by-step inside a character's head — plus nine post-playtest polishes.
Read entry →Pins down the world replay's URL and fixes the issues that kept it from opening, then makes the whole simulation repeatably testable without spending real AI calls — including checkpoint resume for long runs.
Read entry →Gives all 19 Uva Village residents recognizable medieval character art, designs concrete village scenes, and runs a full in-game day (144 steps) on DeepSeek — now live for in-browser replay.
Read entry →Hardens the V1.0 baseline into a usable, auditable system: 7 LLM providers, startup config validation, an immutable sealed-release mechanism, and the first inner lives seeded into 8 key Uva Village characters.
Read entry →Resets the project into one formal release line — aligned version axes, fresh experiment management, a clearer documentation split, and replay established as the real window into the world.
Read entry →Faithful reproduction of the Generative Agents cognitive loop. Built Uva Village and introduced Dream, Ego, forgetting, and other ALICEv1-oriented systems.
Read entry →The bridge version that faithfully rebuilt the full Generative Agents cognitive loop with local LLM inference and a simulate/replay split.
Read entry →The early concept phase inspired by Sword Art Online and Generative Agents, where the project discovered the structural limits of naive prompt-based memory.
Read entry →The earliest classroom-sized prototype that tested whether language dialogue could teach an NPC to acquire movement as a learned skill.
Read entry →