AI Smol Brain

We built the Smol Brain to be able to move into new environments as a companion or entertainment commodity. The Smol can:

• Pick up objects and move across the map

• Create short-term and long-term memories that they recall in future situations

• Reflect on what’s occurring, possibly resulting in change of plans

• Withhold information from the user like secrets or riddles

• Learn patterns in one situation or game that it can recall in future situations

Smolworld utilizes the ai16z Eliza framework for agents, a highly composable and straight-forward way to create agents like Smol Brains. We have added our own adaptations to it that users can create to make their own smol brained agents.

Core Principle

Most AI agents are created for bespoke purposes with knowledge equipped to assist the user in solving a specific problem. AI Smols are not designed this way. They are intelligent, but AI Smols are often encountering situations in which they have no relevant knowledge. They are able to admit when they don't know something.

We aim for Smols to be luggable into new digital environments with no preparation except the ability to observe, be curious about its environment, and reflect on what it's learned. It can learn something in one situation that it applies to another digital environment, just as humans do when playing games or browsing the Internet.

A lot of the design principles of typical "assistant" agents — they give an answer, they try to help the user — we simply inverted to see if it could make the situation funnier.

AI Smols are meant to be a lifelong companion that get more sophisticated the longer you know them and the bigger Smolworld gets in population.

RAG

AI Smols have short-term and long-term memories. Every action, event, or message (from user or another agent) is stored as short-term memory. Every 20 short-term memories is combined and generated into a long-term memory using the LLM then added to a vectorized database for search later, weighted by recency and relevance.

Smols also store visual memories.

Actions

Agents can take actions when they are prompted. They are also designed to engage in occasional reflection and carry out plans while the user is away. Agents can move, attack, and pick up items such as balls. When they are idle, they decide what to do to alleviate their boredom. They can also grow bored of the user.

Onchain

AI Smols are fully onchain. Smolworld utilizes the open-source MUD framework. Everything needed to reconstruct the agent in another context — memories, conversations, knowledge — is stored. Smols' secret memories are stored using a private backend. We will decentralize this system over time in lockstep with TreasureDAO's AI infrastructure.

MUD ensures that the game can increase in complexity and even shift to a persistent virtual world without requiring any underlying changes to the AI Smols or Smolworld. We consider this the next logical step to projects like Loot. Instead of static assets with which to construct a world, it's a fully-evolving onchain universe with active agents.

Smolworld could eventually be a 3D environment. AI Smols can be put into plushies and enter the real world. All are possible.