- Function: Primary AI component that interacts with the external environment, termed as the "world".
- Interactions: Processes user queries, observes real-world events, analyzes data streams, etc.
- Output: Generates structured data points with context, termed as "memories".
- The external environment or context in which the LLM Agent operates.
- Definition: Structured data points with context generated from the LLM Agent's interactions with the world.
- Perspective: Stored in the first person, from the viewpoint of the LLM Agent.
- Metadata: Contains information about the experience, such as whether it was positive, negative, expected, unexpected, etc.
- Purpose: Serve as the foundational data for refining the agent's understanding and responses, and act as a feedback mechanism.
- Type: Specialized database designed for AI workloads.
- Function: Efficiently handles high-dimensional data, making it ideal for storing and retrieving memories.
- Process: Periodic or overnight refinement of memories.
- Rationale: Batching the refinement process optimizes computational resources and allows for efficient use of QLoRA finetuning outside peak operational hours.
- Function: Further refines memories during the consolidation process using the metadata as a form of reinforcement signal.
- Output: Produces the LLM Adapter.
- Definition: A refined knowledge module resulting from the consolidation and QLoRA finetuning processes.
- Integration: When integrated back into the LLM Agent, it enhances the agent's decision-making, accuracy, and predictive capabilities.
With the inclusion of first-person memories and metadata-driven reinforcement learning, the system is designed to allow the AI agent to learn and evolve in a manner that mimics human experiential learning. This approach aims to make the agent's interactions more context-aware and adaptive.