Build Memory from Scratch

Framework abstractions hide failure modes. Building once teaches you exactly what gets stored, how retrieval works, and where latency hides.

Building memory from scratch is like writing your own ORM once — painful, but you never misunderstand what the database is doing.

Scenario

A 100-line Python module that remembers user preferences across Streamlit sessions using Chroma and OpenAI embeddings.

Solution

In Agent Memory, apply Build Memory from Scratch to this scenario: A 100-line Python module that remembers user preferences across Streamlit sessions using Chroma and OpenAI embeddings. 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 Build Memory from Scratch (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 Build Memory from Scratch happens in the code.

1def update_memory(user_id, conversation, llm, store):  # define a reusable function2    facts = llm.invoke(f"Extract new facts as bullet points:\n{conversation}").content3    for fact in facts.split("\n"):4        if fact.strip():5            store.add(texts=[fact], metadatas=[{"user_id": user_id}], ids=[str(uuid4())])6
7def inject_memory(user_id, query, store):  # define a reusable function8    hits = store.similarity_search(query, k=5, filter={"user_id": user_id})9    return "\n".join(h.page_content for h in hits)  # return the result

• Treating Build Memory from Scratch as a black box without evaluation
• Ignoring cost and latency in production
• Skipping error handling for build memory from scratch

Quick recap — the most important points from this module.