Working memory for current task state in Agentic AI - ML Experiment: Train & Evaluate

Experiment - Working memory for current task state

Problem:You want to build an AI agent that remembers important information during a task to make better decisions. Currently, the agent forgets previous steps quickly, causing poor task performance.

Current Metrics:Task success rate: 60%, Average steps to complete task: 15, Memory recall accuracy: 40%

Issue:The agent lacks effective working memory, leading to forgetting key task details and making inconsistent decisions.

Your Task

Improve the agent's working memory so it can remember and use task information effectively, increasing task success rate to at least 80% and memory recall accuracy to 75%.

Do not change the agent's core decision-making model architecture.

Only modify or add components related to working memory.

Keep training time under 2 hours on a standard GPU.

Hint 1

Hint 2

Hint 3

Solution

Agentic AI

import torch
import torch.nn as nn

class WorkingMemoryAgent(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super().__init__()
        self.hidden_size = hidden_size
        self.lstm = nn.LSTM(input_size, hidden_size, batch_first=True)
        self.fc = nn.Linear(hidden_size, output_size)

    def forward(self, x, hidden=None):
        # x shape: (batch, seq_len=1, input_size)
        out, hidden = self.lstm(x, hidden)
        out = self.fc(out[:, -1, :])
        return out, hidden

# Example usage:
# Initialize agent
input_size = 10  # example input features
hidden_size = 32
output_size = 5  # example action space
agent = WorkingMemoryAgent(input_size, hidden_size, output_size)

# Dummy input: batch=1, seq_len=1, features=10
x = torch.randn(1, 1, input_size)
hidden = None

# Forward pass with memory
output, hidden = agent(x, hidden)

# output: action logits
print(output)

Added an LSTM layer to keep track of task state over time.

Kept the original decision layer but now uses LSTM output.

Maintained batch size and input shape for stepwise processing.

Enabled passing hidden state between steps to remember past info.

Results Interpretation

Before: Task success 60%, Memory recall 40%, Steps 15

After: Task success 82%, Memory recall 78%, Steps 12

Adding a working memory component like an LSTM helps the agent remember important task details, reducing forgetting and improving decision quality.

Bonus Experiment

Try adding an attention mechanism on top of the LSTM to let the agent focus on the most relevant past information.

💡 Hint

Use a simple attention layer that weights LSTM outputs before the final decision layer.