Working Memory

Working Memory Agent for the Arshai framework.

This module provides an agent specialized in managing conversation working memory. It handles fetching conversation history, updating memory in storage systems, and generating memory summaries.

Classes

class arshai.agents.hub.working_memory.WorkingMemoryAgent(llm_client, system_prompt=None, memory_manager=None, chat_history_client=None, **kwargs)[source]

Bases: BaseAgent

Agent specialized in managing conversation working memory.

This agent is responsible for maintaining and updating the working memory of conversations. It fetches conversation history, generates summaries, and stores updated memory in persistent storage.

Capabilities: - Fetches conversation history from chat history storage - Retrieves current working memory from storage (e.g., Redis) - Generates revised working memory based on new interactions - Stores updated memory for future reference

Requirements: - Memory manager for storage operations - Chat history client for conversation retrieval (optional)

Example

memory_manager = RedisMemoryManager(redis_client) agent = WorkingMemoryAgent(

llm_client, “You are a memory management assistant”, memory_manager=memory_manager

)

# Update memory for a conversation result = await agent.process(IAgentInput(

message=”User asked about product pricing”, metadata={“conversation_id”: “123”}

)) # Returns “success” if successful, or “error: <description>” if failed

__init__(llm_client, system_prompt=None, memory_manager=None, chat_history_client=None, **kwargs)[source]

Initialize the working memory agent.

Parameters:

  • llm_client (ILLM) – The LLM client for generating memory updates

  • system_prompt (str) – Optional custom prompt (uses default if not provided)

  • memory_manager (IMemoryManager) – Memory manager for storage operations

  • chat_history_client (Any) – Optional client for fetching conversation history

  • **kwargs – Additional configuration

async process(input)[source]

Process memory update request.

This method: 1. Extracts conversation_id from metadata 2. Fetches current memory and history 3. Generates updated memory using LLM 4. Stores the updated memory

Parameters:

input (IAgentInput) – Input containing the new interaction and metadata with conversation_id

Returns:

Status of the operation - “success” if successful,

”error: <description>” if failed, or None if no conversation_id

Return type:

str