Game Features: Remnant

Game Remnant (games/generated/Remnant/) is a generated variant of AgentOdyssey that we use for Experiment 1 of the paper. It is built on top of the base game and designed to create a significantly more dynamic and challenging text-world environment. Where the base game establishes a clean, predictable sandbox with 17 action rules and 9 step rules, Remnant expands the action vocabulary to 22 rules and the step-rule roster to 18 rules, adding entire new subsystems around stealth, survival, environmental pressure, and time-of-day dynamics.

This page documents every difference from the base game. Mechanics that are identical to the base game are not repeated here; see Features: Base for those.

Motivation of Remnant

The core idea of Remnant is to generate a readily-playable game with significantly more mechanical depth and complexity than the base game, while still being simple enough to be learnable by an LLM agent within a reasonable number of steps. It shall also include mechanisms that better test the key capabilities of successful test-time continual learning agents as described in the paper: episodic memory, skill learning, exploration, world knowledge acquisition, and long-horizon planning.

World Definition

Remnant ships a much larger world definition compared to the base game's compact starter map. The headline numbers: 7 places and 17 areas (up from 2 / 6), 13 NPC types including 5 specialized merchants and 4 enemy variants plus a unique boss (up from 1 merchant, 1 scout, 1 enemy), and roughly double the object catalog with new raw materials, food items, tools, weapons, armor, and containers spread across the new regions. Each enemy type defines an explicit combat pattern in the JSON, and the world parser auto-assigns a default pattern to any enemy that lacks one.

The crafting tree is both wider and flatter. Remnant adds new material chains rooted in the new regions, such as field grain that milled into flour, while simultaneously removing the nugget-stage intermediates (stone chunk, iron nugget, star metal nugget) so that ore smelts directly into bars.

Lighting

Remnant introduces a visibility system tied to area lighting. Each area in the world definition carries a light flag (in the base game this flag exists but is randomized and unused). When an agent is in a dark area and is not holding a torch, the observation is reduced to just the location name, current time, and the agent's own stats. The agent cannot see objects on the ground, NPCs in the area, or neighboring paths. Holding a torch (or a lantern, which shares the light usage tag) in hand restores full visibility. This adds a resource-management layer: torches occupy a hand slot that could hold a weapon or item, and the agent must remember which areas are dark and plan accordingly.

New Action Rules

Remnant introduces five new actions that the agent can perform. These actions are not taught in the tutorial. After the tutorial completes, the environment lists these extra verbs (without descriptions) so the agent knows they exist.

eat <object> — Consume a food item to restore HP. The rule searches for the named item across the agent's hands, inventory, held containers, and the ground (in that priority order) and removes one unit from wherever it is found. The item must either belong to the food category, have a consume usage tag, or carry a nonzero hp_restore attribute.

throw <object> <npc> — Hurl a hand-held object at an NPC in the same area. The thrown object deals instant damage equal to its attack_power attribute (minimum 1) without triggering an immediate counterattack from the target. After impact, the object has a 70% chance of landing intact on the ground in the current area and a 30% chance of shattering. When an object shatters, its crafting ingredients are partially salvaged and deposited on the ground using the same yield formula as disassemble (see below). Containers and writable objects cannot be thrown. If the NPC's HP drops to zero from the throw, the kill is resolved identically to a standard attack kill: loot drops, kill counters update, and events are emitted.

lockpick <area> — Attempt to silently unlock a locked neighboring path without using a key. The agent must be holding a lockpick tool or a tension wrench. A lockpick gives a 20% success chance; a tension wrench gives 10%. On success, the path is unlocked in both directions (just like using a key) and one unit of the tool is consumed.

disassemble <object> — Break down a craftable object into some of its original ingredients. The target must have a crafting recipe (non-empty craft_ingredients), must be either in the agent's hand or on the ground in the current area, and must not be currently equipped. The salvage yield follows a deterministic formula per ingredient: if the recipe calls for 2 or more of that ingredient, the agent recovers floor(50% × required_amount); if it calls for exactly 1, there is a 50% random chance of recovering 1 unit; quantities below 1 yield nothing. Recovered materials are placed on the ground.

pickpocket <npc> — Attempt to stealthily steal from an NPC in the current area. The base success chance is 40%, modified by circumstances: empty hands grant +10%, having two or more items in hand imposes -20%, and targeting a merchant costs an additional -15%. On success, the agent steals either coins (1 to 5, taken directly from the NPC's coin pool) or one random item from the NPC's inventory. Stolen goods go into the agent's hand if a hand slot is free, otherwise they drop on the ground. On failure, the NPC retaliates and the agent takes damage equal to the NPC's attack power minus the agent's defense (minimum 1).

New Step Rules

Remnant adds ten new step rules that create an interconnected simulation layer on top of the base game. Many of these rules interact with each other through shared state like the noise system and the step-modifier dictionary. Most of these rules are devised and implemented entirely by the rule generator without human descriptions.

Day Cycle (priority 4) The day cycle divides time into a 24-hour loop at a rate of 6 steps per hour. Seven distinct periods rotate through the day, each applying different combat modifiers via the _step_modifiers system:

Period	Hours	Ambush Modifier	Enemy Damage
Dangerous Night	0:00–1:00	+30% additive	×1.30
Late Night	1:00–6:00	—	×1.20
Dawn	6:00–7:00	—	—
Morning	7:00–12:00	—	—
Peaceful Midday	12:00–13:00	Capped at 5%	×0.80
Afternoon Calm	13:00–18:00	—	×0.80
Evening Inspiration	18:00–24:00	—	—

Beyond combat modifiers, two periods have special effects. At dawn, areas that contain a merchant NPC receive a small coin deposit (1 to 3 coins) on the ground. During evening inspiration, each crafting event that occurred on that step has a 25% chance of producing a bonus copy of the crafted item, rewarding agents who time their crafting activity.

When a period transition occurs, agents receive atmospheric feedback messages (e.g. "Tension tightens across the world" for dangerous night, "Edges soften" for peaceful midday). These are purely informational but serve as temporal cues that a memory-capable agent can learn to associate with mechanical changes.

Wandering Patrols (priority 5) Non-unique merchants and enemies may wander to a neighboring unlocked area each step with a 6.25% probability. The rule skips unique NPCs, quest NPCs, NPCs currently in active combat, and any NPC in the tutorial area. Movement is also capped so that the number of non-unique NPCs in the destination area does not exceed the area's level.

When a merchant wanders into a new area, it restocks 1 to 2 random items in its inventory and agents present in the destination area see a message that the merchant has arrived and refreshed stock. When an enemy wanders in, agents in that area are warned, and the ambush_prob_additive modifier is increased by 0.05 for the current step, making the area more dangerous.

Wandering patrols mean that the NPC landscape is no longer static. Merchants may leave an area the agent was counting on, and enemies may appear in previously safe zones. This creates a memory challenge: the agent needs to track where NPCs have moved rather than relying on a fixed world map.

Enemy Respawn (priority 8) Every 25 to 45 steps, the rule checks for areas that currently contain no enemy NPCs (excluding the spawn area and the tutorial area). It then spawns new non-unique enemies in 1 to 3 of those empty areas, chosen randomly. The new enemies are instances of existing enemy prototypes from the world definition, leveled to match the area they are placed in. Agents in the affected area are warned that a hostile creature has appeared.

This prevents the world from becoming completely safe after the agent has cleared all enemies, ensuring that combat remains a persistent concern throughout the game.

Continuous Crafting Momentum (priority 5) When an agent crafts in the same area on consecutive steps (within a 5-step window), a momentum streak is tracked. Starting from the second consecutive craft in the same location, the agent receives a coin bonus: 1 to 2 coins per crafted item materialize on the ground. This encourages agents to plan crafting sessions by gathering all materials in one location and crafting in quick succession rather than scattering crafting across different areas and steps.

Forgotten Caches (priority 4) Areas that no agent has visited accumulate an idle counter each step. When the counter crosses a threshold (randomly set between 35 and 50 steps), there is a 70% chance that a small resource cache spawns on the ground. The cache contains 1 to 3 items drawn from materials and currency objects that are normally distributed to that area. Higher-level areas (level 3+) get 2 to 3 items instead of 1 to 3. If the area contains a merchant, there is an additional 25% chance for one extra tool or material. The presence of enemies in the area halves the accumulation speed, and having agents in the area resets the counter entirely.

This creates a reward for revisiting previously explored but long-abandoned areas, testing the agent's ability to remember the map layout and plan return trips.

Soundscape Pressure (priority 5) Every area tracks a floating-point noise level that rises from in-game events and decays over time. Different event types contribute different noise weights: combat events like kills and damage add 1.0 to 1.5 noise, crafting and trading add 0.5 to 0.8, item manipulation adds 0.3, and so on. The noise level is capped at 12.0 and decays each step.

Areas become mechanically different based on their noise level:

• Loud areas (noise ≥ 8.0): The ambush_prob_additive modifier increases by 0.05 for agents in the area, and there is a 3% chance per step that a weakened scout enemy spawns. Scouts have 50% of normal HP and 30% of normal attack power.
• Hushed areas (noise ≈ 0 for 4+ consecutive steps): The ambush_prob_additive decreases by 0.05, and there is an 18% chance per step that a small cache (1 coin or 1 low-tier material) appears.

The agent can actively manage noise by writing "quiet" on a piece of paper and dropping it in an area. This instantly zeroes the area's noise level and the feedback reads "The dropped note whispers 'quiet' and calms the area." This mechanic rewards agents that can remember and apply the dampening trick at the right time.

Rumor Mill (priority 5) After notable events (NPC kills, merchant arrivals, path unlocks, quest stage advances), areas that contain a merchant have a 12% chance of spawning a scribbled rumor note on the ground. Rumor notes are writable instances that contain hint text, randomly selected from a pool of tips about: time-of-day effects, merchant stock, area and place names, and main quest NPC locations.

Picking up and inspecting a rumor note while holding it awards 1 to 2 coins (one-time reward per note). Dropping a rumor note in an area different from where it originated reduces that area's noise by 0.8 and lowers ambush probability by 0.02, acting as a portable noise dampener. Rumor notes expire after 6 to 12 steps and are automatically removed from the ground if not picked up.

Shifting Paths (priority 6) Every 6 to 12 steps, the rule may temporarily lock or unlock 1 to 2 paths in the world. The probability of a lock or unlock event on a given path is influenced by two factors:

• Noise: Loud areas increase lock probability by 50% and decrease unlock probability. Hushed areas increase unlock probability by 50% and reduce lock probability.
• Time of day: Night periods increase lock probability by 30%; daytime and calm periods increase unlock probability by 30%.

Temporary shifts last 3 to 6 steps before automatically reverting to their original state. Safety checks ensure that the rule never locks the last open exit of an area that currently has an agent inside, and it never affects the spawn area, tutorial area, or quest-gated paths.

Agents in adjacent areas receive feedback about paths slamming shut or opening briefly, giving them a window to plan movement around the temporary change.

Carrion Scavengers (priority 5) Areas where combat has occurred or where uncollected loot sits on the ground accumulate scavenger interest. When interest is high enough, scavengers may convert low-value ground items into coins at a 50% depreciation rate (e.g. an item worth 4 coins becomes 2 coins). The number of items scavenged depends on time and environment: during the night or in loud areas, scavengers take up to 3 items; during dawn or when agents are present, they take at most 1. Each scavenge event also adds 0.35 noise to the area and increases the ambient ambush probability by 0.02.

This creates urgency around loot collection: if the agent leaves valuable items on the ground after a fight, scavengers will gradually convert them into less-valuable coins and raise the area's danger level.

Food Respawn (priority 4) Every 25 steps, all food-category objects respawn in the areas where they are normally distributed. Each food type can spawn 0 to 4 units per area. Agents present in an affected area are notified about the fresh food appearing. Combined with the eat action, this ensures that health recovery remains available throughout long game sessions.

Changed Step Rules

Active Attack: The base game's Active Attack rule scales ambient attack probability from 5% to 15% across area levels. Remnant raises the maximum to 25% to account for higher maximum NPC levels and adds support for the _step_modifiers dictionary that other rules write to. Before computing the final attack probability, the rule checks for five modifier keys:

• ambush_prob_override — If set, replaces the base probability entirely.
• ambush_prob_additive — Added to the base probability.
• ambush_prob_multiplier — Multiplies the final probability.
• enemy_damage_multiplier — Scales all ambient attack damage.
• enemy_attack_damage_bonus — Flat damage bonus added to ambient attacks.

This modifier system is the central integration point for Remnant's new step rules. The Day Cycle sets baseline modifiers each period, Wandering Patrols and Soundscape Pressure add incremental adjustments, and Carrion Scavengers and Rumor Mill contribute their own tweaks.

System Interactions

One of Remnant's defining characteristics is how its new rules form an interconnected web of cause and effect. Understanding these interactions is important for appreciating the complexity of the environment:

• Combat → Noise → Scouts → More Combat: Killing enemies raises area noise, which can trigger scout spawns from the Soundscape Pressure system, which in turn raises noise further.
• Noise → Shifting Paths: High noise makes paths more likely to temporarily lock, potentially trapping agents or blocking escape routes.
• Time of Day → Everything: The Day Cycle's modifiers affect ambush probability, enemy damage, crafting bonuses, path shifting bias, and scavenger behavior.
• Wandering Patrols → Ambush Probability: Enemy arrival increases the step modifier, making the area immediately more dangerous.
• Inaction → Forgotten Caches: Leaving areas unvisited spawns resources, rewarding agents with good spatial memory.
• Noise Dampening: Writing "quiet" on paper and dropping it, or transporting rumor notes to different areas, actively reduces noise and ambush probability.
• Crafting Momentum → Economy: Consecutive crafting in one location generates bonus coins, which feed into trading.
• Scavengers → Economy + Noise: Uncollected loot is converted to coins at a loss and raises noise, penalizing agents that don't collect items promptly.

These interlocking systems mean that Remnant heavily rewards agents with strong long-term memory. Remembering the day cycle timing, tracking NPC movements, recalling which areas are noisy, and planning efficient crafting sessions all depend on the agent's ability to retain and use information from many steps in the past.