Managing Agent Memory

AI Agents with Hugging Face smolagents

Adel Nehme

VP of AI Curriculum, DataCamp

Stateless by Default

Each .run() call is a fresh start.

Example:

career_advisor.run("What career skills should I highlight?")

You should highlight Python, SQL, data visualization, 
machine learning fundamentals, and communication skills tailored to business outcomes.

career_advisor.run("Can you format those skills as bullet points?")

Sorry, I'm not sure which skills you're referring to. Could you clarify?

Retaining Memory Between Interactions

career_advisor.run("What career skills should I highlight?")

You should highlight Python, SQL, data visualization, 
machine learning fundamentals, and communication skills tailored to business outcomes.

Pass reset=False:

career_advisor.run("Can you format those skills as bullet points?", reset=False)

Sure! Here are the skills as bullet points:
- Python
- SQL
- Data visualization
...

Memory Helps You Debug, Too

User: What's the expected salary?

Agent: It's $80,000

User: Wait, that seems wrong...

Agent: Sorry, I'm not sure what you mean

Inspect what happened in the agent's run:

Reviewing all code the agent generated
Tracing its reasoning, actions, and tool usage
Debugging incorrect answers or broken logic

What Code Did the Agent Run?

The .return_full_code() method lets you see all executed code.

executed_code = career_advisor.memory.return_full_code()
print(executed_code)

# ...other steps omitted for brevity

salary = 80000  # <- hardcoded?

# script continues...

What Was the Agent Thinking?

conversation_steps = career_advisor.memory.get_succinct_steps()
print(conversation_steps[5])

{
  "step_number": 5,
  "tool_calls": [
    {"function": {"name": "python_interpreter"}},
    {"function": {"name": "web_search"}}
  ],
  "code_action": "import requests\nskills = requests.get('api.jobsearch.com').json()",
  "observations": "resume_agent found 15 relevant skills for transition",
  "token_usage": {"total_tokens": 334},
  ...
}

Save Agent Sessions for Analysis

import json

def save_agent_memory(agent):
    with open("agent_memory.json", "w") as f:
        json.dump(agent.memory.get_succinct_steps(), f, indent=2, default=str)

# Save memory to a file
save_agent_memory(career_advisor)

Logs can help with:

Post-hoc analysis
Regression testing
Improving agent behavior over time

Fixing Agent Failures: What to Adjust

Memory issues: Use reset=False or reset intentionally
Reasoning problems: Try a stronger model
Inconsistent behavior: Improve system prompt
Tool confusion: Clarify tool docstrings

Let's practice!

AI Agents with Hugging Face smolagents