For agents that use memory, task success rate and long-term consistency are key metrics. Memory helps agents remember past actions and information, so they can make better decisions over time. Measuring how often the agent completes tasks correctly (success rate) and how well it keeps consistent behavior across steps (consistency) shows if memory is helping.
Task Completion Confusion Matrix:
| Predicted Success | Predicted Failure
------|-------------------|-----------------
Actual Success | 85 (TP) | 15 (FN)
Actual Failure | 10 (FP) | 90 (TN)
Total tasks = 200
Precision = TP / (TP + FP) = 85 / (85 + 10) = 0.894
Recall = TP / (TP + FN) = 85 / (85 + 15) = 0.85
F1 Score = 2 * (Precision * Recall) / (Precision + Recall) ≈ 0.871
This matrix shows how well the agent with memory predicts task success. High precision means it rarely says success when it fails. High recall means it catches most successes.
Imagine an agent helping a user book flights. If it has high precision, it rarely suggests wrong flights (few false positives), so the user trusts it. But if it has low recall, it might miss some good flight options.
If it has high recall, it finds almost all good flights, but with low precision, it might suggest many bad options, annoying the user.
Memory helps balance this by remembering past preferences and avoiding repeated mistakes, improving both precision and recall over time.
Good metrics: Task success rate above 85%, precision and recall both above 80%, and consistent behavior across sessions.
Bad metrics: Success rate below 60%, precision or recall below 50%, and erratic or contradictory actions showing poor memory use.
Good memory use means the agent learns from past steps and improves. Bad memory use means it forgets or repeats errors.
Accuracy paradox: An agent might have high overall accuracy by guessing common outcomes but fail on important rare tasks.
Data leakage: If the agent's memory accidentally includes future information, metrics look better but don't reflect real use.
Overfitting: The agent might memorize specific past tasks perfectly but fail to generalize to new ones, showing high training success but low real-world performance.
Your agent has 98% accuracy but only 12% recall on important tasks. Is it good for production? Why not?
Answer: No, it is not good. The low recall means the agent misses most important tasks, even if overall accuracy is high. This means it often fails when it matters most, so memory or decision-making needs improvement.