Memory Expiration Strategies: Cleaning the Mind

A vector database that grows forever becomes a liability.

1. Noise: Old, irrelevant memories make it harder to find the "Signal."
2. Cost: You pay for every GB of vector storage.
3. Latency: Searching through millions of "Hi" and "How are you?" messages slows down every request.

In this lesson, we learn how to "Prune" the agent's mind using Expiration Strategies.

---

1. Time-to-Live (TTL): The Standard Pruning

The simplest way is to delete memories that are too old.

• Application: "If a conversation happened more than 90 days ago, it is likely obsolete."

The Meta-Filtering approach: Instead of deleting, you can simply Filter queries: where={"timestamp": {"$gt": ...}}. Periodically, a background job can delete the records that are no longer reachable by the filter.

---

2. Importance-Based Retention

Not all memories are created equal.

• "What is your name?" -> Low Importance.
• "I am allergic to peanuts." -> High Importance.

The Score Metric: Have the LLM assign an importance_score (1-10) during storage. The Policy: "Only delete items with an importance score less than 3 after 30 days. Keep items greater than 8 forever."

---

3. The "Decay" Function

A sophisticated agent uses a Decay Score influenced by both Recency and Frequency.

Final_Score = Base_Importance * (0.5 ^ Years_Passed) * (Number_of_Retrievals)

If a memory is old but it is "Recalled" every day, it resets its decay timer. This is how humans preserve core identity while forgetting yesterday's lunch.

---

4. Implementation: The Cleanup Worker (Python)

Python Code: Periodic Memory Pruning

def prune_agent_memory(collection, importance_threshold=2, days_old=30):
    cutoff_timestamp = time.time() - (days_old * 86400)
    
    # Identify candidates for deletion
    results = collection.get(
        where={
            "$and": [
                {"timestamp": {"$lt": cutoff_timestamp}},
                {"importance": {"$lt": importance_threshold}}
            ]
        }
    )
    
    if results['ids']:
        collection.delete(ids=results['ids'])
        print(f"Pruned {len(results['ids'])} obsolete memories.")

---

5. Summary and Key Takeaways

1. Avoid Infinite Growth: Set hard policies for data retention.
2. Importance Metadata: Scoring memories at the point of ingestion is mandatory for scaling.
3. Filter Before Deleting: Use queries to "Soft-Delete" first to ensure system stability.
4. Decay Logic: Reward "Frequency of access" to keep important old knowledge alive.

---

Exercise: The Memory Auditor

1. Look at a log of 100 interactions.
2. Identify 5 memories that you think the agent should keep "Forever."
3. Identify 20 memories that should be deleted after 24 hours.
4. The Question: What single word or piece of metadata would allow an automated script to distinguish between these two groups?

---

Congratulations on completing Module 12! You have mastered the psychology of agentic vector storage.