How to Make an AI Agent Remember: Memory vs Playbooks

AI agent memory holds single rules and facts the agent applies wherever they fit, such as "net sales excludes shipping". A playbook holds the full method for one kind of job, such as the monthly refund report. Use memory for what your agent should always know and playbooks for what it should do step by step.

Without either, you repeat yourself forever. You correct the agent on Monday, and on Tuesday it makes the same mistake. This guide explains the two ways to fix that, what to teach first, and how to keep control of what the agent has learned. The examples use Concierge, and the ideas carry over to other agents.

A quick test: if you can say it in one sentence and it applies across many jobs, it is a memory. If you would have to walk a new hire through it, it is a playbook.

In Concierge, memories come from four places:

• Telling an agent: Say "From now on, put refunds first in the weekly report." It saves that right away and shows "Saved to memory" with Undo and Edit.
• Handoffs: After your team handles a conversation the agent handed over, it suggests what it learned, such as how the team resolves that kind of request.
• Work that went well: After a job that used tools succeeds, the agent writes a short journal of the steps that worked and uses it on similar work.
• Writing it yourself: Add one by hand on the Memory page.

What your team does matters most here, because learning from handoffs turns normal work into training. You don't need to set aside time to teach. You correct the agent as you would a new hire, and the correction stays.

Lessons customers will see need a little more proof. By default, an answer your team gave goes live right away, and other suggestions go live once handoffs from two different customers back them up. If you'd rather approve everything yourself, one setting makes every customer-facing suggestion wait for a person.

• The job has several steps, and the order matters.
• A number could be calculated more than one way, and you need the same answer each time.
• The output has a fixed format, such as the columns in a report.
• Different people ask for the same job and should get the same result.
• You keep making the same three corrections to the same job.

That last sign is useful: when memories start to pile up around one job, combine them into a playbook. In Concierge, you can also open the Journals view and use Turn into playbook to start a draft from an approach that already worked. Our guide on how to write a playbook includes a template.

• Your definitions: What do active subscriber, net sales and late order mean at your company? These cause the most confusion.
• Your policy exceptions: For example, "We ship replacements for damaged items without waiting for a return."
• Your report preferences: Which number comes first, which time zone you use and how many rows you want.
• Your tone: Words you always use, and words you never want a customer to read.
• Your most repeated job: Write one playbook for the job you ask for most often.

Don't try to write everything on day one. Teach what comes up. A memory you added because of a real mistake is far more useful than twenty you guessed at.

A team preference shouldn't reach customers, and a personal one shouldn't reach your team. Each Concierge memory has one audience:

• Customer-facing: Used with customers and teammates. Handoff lessons are always customer-facing.
• Team: Used only with teammates, for things like report formats.
• Only me: Used only in your own private conversations with your agents.

These are two separate kinds of memory, and it is worth keeping them apart. Customer memory is the history of one customer across channels, so nobody has to repeat their story. Learned memory is how your business works, and it never names a customer.

Concierge enforces that line. Before a model reads a handoff, the customer, their orders and the teammates involved are replaced with placeholders. Anything learned automatically is checked and refused if it contains an email address, phone number, order number, tracking number or a link with a token, or if it describes a specific person or case. If you tell an agent something about one customer, it won't go into memory. It tells you to save it as a note on that customer instead.

Anything an agent learns should be visible and removable. In Concierge these controls live on the Memory page:

Stale memories are handled too. A memory nobody has confirmed in 180 days shows "Not in use" and is skipped. Editing or pinning it puts it back in use, and each time a successful job applies a memory, that counts as confirmation.

• Pinning everything: Agents use a limited number of memories per request. Pin only what applies to every request.
• Burying a method in memories: Ten memories about one report should be one playbook.
• Vague rules: "Be careful with refunds" gives the agent nothing to act on. "Ask for a photo of damaged items" does.
• Expecting a playbook to add powers: A playbook guides how an agent works. It cannot give the agent capabilities it doesn't have.
• Never reviewing: Policies change. Read through your memories when they do.

The pattern is simple: say what should happen next time, in words that would still make sense to a new hire who never saw the original mistake.

Concierge learns from how your team works, follows the playbooks you write and keeps each customer's history across channels. Every workspace starts with analysis playbooks you can edit. The steps are in the memory and learning docs and the playbook docs.

Keep reading: how to write a playbook, how to check an agent's work and where your data goes. Glossary: AI memory, AI playbook and standard operating procedure.

Some do, and what they keep varies. A useful agent remembers rules about how your business works and the history of each customer, and it lets you see, edit and erase what it has learned. Ask each vendor what is kept and who can see it.

A memory is one rule or fact the agent applies wherever it fits. A playbook is a full method for one kind of job, with steps, definitions and an output format. Use a memory for a single correction and a playbook for a repeatable job.

No. Learned memories describe how your business works and never name a customer. Automatic learning is refused if it contains details like an email address, phone number or order number. Customer history is kept separately, so customers don't have to repeat themselves.

In Concierge, click Forget on the Memory page to erase a memory's text for good. You can also ask an agent in chat to forget one of your own memories or a team memory. Archive is the reversible choice.