Long-term memory

[mdw771]

Overview

Support long-term agent memory through RAG.

Each time the user adds a message, a rule-based detector checks if the message is notable. If so, it is embedded and stored in the vector DB. For RAG, the vector DB is searched and matched records are prepended to the context.

Usage

With the task manager, enable RAG through memory_config:

task_manager = BaseTaskManager(
    llm_config=llm_config,
    memory_config=MemoryManagerConfig(
        enabled=True,
        write_enabled=True,
        retrieval_enabled=True,
        top_k=5,
        score_threshold=0.25,
        min_content_length=12,
        embedding_model="text-embedding-3-small",
        vector_store_path="memory.db",
    )
)