Memory Ranking

Top-k vector search returns plausible but not always best results. Ranking ensures the most useful memories win the limited token budget.

Memory ranking is a search engine's second page — vector search is the crawl; ranking is the algorithm that puts the best answer first.

Scenario

Five memories match 'deployment.' Ranking boosts the one from yesterday about the staging deploy over a six-month-old note about deployment theory.

Solution

In Agent Memory, apply Memory Ranking to this scenario: Five memories match 'deployment. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Memory Ranking (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Memory Ranking happens in the code.

1# Memory Ranking — minimal example2from openai import OpenAI3
4client = OpenAI()  # create API client5
6# Ask the model to explain this topic7response = client.chat.completions.create(  # core API call for Memory Ranking8    model="gpt-4o-mini",9    messages=[10        {"role": "system", "content": "You explain memory ranking clearly."},11        {"role": "user", "content": f"What is memory ranking?"},12    ],13    temperature=0,14)15print(response.choices[0].message.content)  # show output for debugging

• Treating Memory Ranking as a black box without evaluation
• Ignoring cost and latency in production
• Skipping error handling for memory ranking

Quick recap — the most important points from this module.