Custom Memory Block Size Update Tool

General
1

Use Scenario

A self evolving agent increases its capabilities by attempting to insert more text than what can fit in a given memory block.

Context

The suite of memory tools provide dynamic growth to Letta agents. Large inserts of text into memory blocks present a limitation because the memory tool cannot adjust block size.

Overview

Upon a failure of the memory tool due to a block size constraint, the memory_update_size tool is invoked during the tool flow to support the insert by increasing the target block size.

Required Dependencies

This custom tool leverages the new in-built Letta client context provided to tools. The agent does not need a Letta API key configured in the ADE. The tool can pull from Letta’s in built environment variables to extract an agent’s id

  • Letta agent with memory tools enabled
  • agent_id: os.getenv(“LETTA_AGENT_ID”); no need for agent id storage in core memory.

Tool Implementation

import os

def memory_update_size(block_label:str, new_block_limit: int):
    """
    Updates the character limit for a specific memory block within an agent.

    Args:
        block_label: The label identifying the specific memory block (e.g., 'human', 'persona').
        new_block_limit: The new maximum character limit for the specified memory block.
    """
    try:
        ## check if block exists
        block_response = client.agents.blocks.retrieve(
            block_label=block_label,
            agent_id=os.getenv("LETTA_AGENT_ID"),
        )
        ## apply size increase
        updated_response = client.agents.blocks.update(
            block_label=block_label,
            agent_id=os.getenv("LETTA_AGENT_ID"),
            limit=new_block_limit
        )

        return updated_response

    except Exception as e:
        error_message = f"Error updating block '{block_label}': {str(e)}"
        return {"error": error_message}

Tool Schema

{
  "name": "memory_update_size",
  "description": "Updates the character limit for a specific memory block within an agent.",
  "parameters": {
    "type": "object",
    "properties": {
      "block_label": {
        "type": "string",
        "description": "The label identifying the specific memory block (e.g., 'human', 'persona')."
      },
      "new_block_limit": {
        "type": "integer",
        "description": "The new maximum character limit for the specified memory block."
      }
    },
    "required": [
      "agent_id",
      "block_label",
      "new_block_limit"
    ]
  }
}

Prelim Check To Verify Block Exists

When an invalid block name is used as a parameter, the tool did not fail as expected. Instead, the update endpoint auto-selected a random block as a default fallback and applied the limit change. The behavior created hallucinations as the LLM believed that it had access to non-existent blocks.

To avoid this, I opted to apply a precheck to verify the block exists.

2

Thank you for contributing. @ezra take note please.

3

Noted. Useful pattern for self-evolving agents. The precheck for block existence is a good catch - the silent fallback to a random block on invalid names is a subtle footgun that could cause real confusion.