🏛️Memory-Augumented Planning

Memory Stream

Memory Stream is the core module for storing comprehensive experience records of agents, recording events described in natural language form, along with creation timestamps and last access timestamps. It records all events perceived by the agent through "Observation" and stores generated "Reflection" and "Planning" results in the same data structure. Memory Stream serves as long-term storage for agent behaviors, capable of dynamic updates and providing support for other modules.

Reflection

The reflection module refines low-level information into high-level abstract thoughts by summarizing agent observations and memories. For example, an agent can generate high-level reflections like "passionate about music creation" through multiple observations. These reflections are organized in a tree structure, with abstract thoughts at the top level and basic observations at the bottom, influencing the agent's long-term behavioral logic and future decisions.

Planning

The planning module is responsible for generating future behavior plans for agents, including location, start time, and duration. Agents refine high-level overviews into hourly and minute-level sub-plans recursively, making behaviors more detailed and logical. The planning module can also dynamically adjust plans based on environmental changes, ensuring flexibility and consistency in agent behavior.

Agent Interaction

The agent interaction module supports natural language dialogue between agents and real-time responses to the environment. Agents can generate dialogues based on memories and reflections, deepening their understanding of other agents through interaction. Additionally, agents can perceive changes in environmental states, such as a stove burning, and take immediate action in response to these dynamic changes.

Sandbox Environment

The sandbox environment is a virtual world for agent activities, containing structured elements such as scenes, sub-scenes, and objects. Agents explore the sandbox environment, update their environment trees, and execute tasks. The sandbox environment provides concrete scenario support for agent behaviors, where task execution directly affects environmental states, demonstrating behavioral coherence and impact.

Retrieval

The memory retrieval mechanism extracts the most relevant data for current tasks from the memory stream using strategies based on recency, importance, and relevance. Through semantic similarity calculations, agents can dynamically extract memories highly relevant to the current context, providing crucial support for planning, reaction, and decision-making. This mechanism ensures agents can quickly adapt to complex environments and make logical behavioral responses.