Agent Memory — How to Make AI Remember Your Customers Across Sessions

Without memory, every agent interaction starts from scratch. With memory, your agent builds knowledge over time — and produces dramatically more relevant, personalised output.

The most common complaint about AI agents after initial excitement wears off is: "It does not remember anything." A customer explains their situation, the agent helps, and the next time the same customer reaches out — the agent knows nothing about them. Every interaction is the first interaction.

This is not an inherent limitation of AI — it is a design choice. Agents can remember. You just need to build the memory system.

Short-Term Memory: The Context Window

Every LLM has a context window — the maximum text it can process in one call. GPT-4o supports 128,000 tokens (roughly 96,000 words). Claude 3.5 supports 200,000 tokens. This is your agent's working memory for a single session — everything in the current interaction.

Managing what goes into the context window is one of the most important intermediate skills. Budget allocation: 30% for the system prompt, 20% for retrieved memory, 40% for conversation history, 10% reserved for the response. Exceeding the window causes silent truncation — often losing the system prompt, which breaks the agent's instructions entirely.

Long-Term Memory with Google Sheets

For most intermediate workflows, you do not need a vector database. A Google Sheet is sufficient for storing structured facts that your agent needs across sessions — customer name, account tier, last issue, preferences, interaction history.

The pattern: at the end of each interaction, write key facts to a Google Sheet row keyed by the customer's email. At the start of each subsequent interaction, look up the row and inject the data into the system prompt. The agent greets the customer by name, already knows their account tier, and continues where the last conversation left off.

== CUSTOMER MEMORY == Customer: {customer_name} | Account: {account_tier} | Last issue: {last_issue_summary} | Preference: {preferred_contact} | Notes: {notes} Use this to personalise every response. Do not ask for information already in the memory.

When to Use a Vector Database

Use a vector database when: you have thousands of customer records that cannot fit in a Google Sheet lookup efficiently, you need to search memory by concept rather than exact key, or you are building a RAG system over your own documents. Pinecone has a generous free tier. Supabase combines relational and vector search. Flowise has a built-in vector store that requires no separate database setup.