Agent Memory

AI memory · LLM memory · AI assistant memory · long-term agent memory

Definition

Agent memory is the storage and reuse of information about a user and past interactions beyond a single request to the model: within the session context (short-term) and in external storage (long-term).

Two kinds of memory

	Short-term	Long-term
Where it lives	In the context window of the current request	In external storage: profile, database, vector index
What it holds	Conversation turns, products shown, filters selected	Size, budget, brands, exclusions, past purchases
Lifetime	One session	Weeks and months — per the retention policy
Example	“Show me the same ones, but in blue”	“Last time you bought a 42 — shall I pick the same size?”

How memory is implemented

History in the context, compressed. The latest turns are passed in full, older ones as a summary, so the window does not overflow.
Structured profile. Facts stored as key–value pairs: shoe size, budget, “no wool”. These are the easiest to verify, show to the shopper and delete.
Vector store. Fragments of past conversations are turned into embeddings and stored in a vector database; on a new request, the system searches for the closest matches by meaning.
Agent notes. Anthropic describes a technique in which the agent regularly writes notes to memory outside the context window and pulls them back in when needed.

The cycle is always the same: the agent writes what matters after a conversation, finds what is relevant before the next answer and inserts it into the context. What to insert, and how much, is a job for context engineering.

Platform memory versus the store’s data

Platform assistants remember the user on their own side: OpenAI began testing memory in ChatGPT in February 2024, and since April 2025 ChatGPT can draw on all past chats. None of this is passed to the store directly. The store’s advantage is its own first-party data: order history, returns, the actual sizes of items bought, the affinity profile. An external assistant has no access to this data — and it is what personal product selection on the site is built on.

In practice: shopping assistant memory without unnecessary risk

Remember what improves selection: sizes, budget, preferences and exclusions, who the purchase is for.
Do not store more than you need: payment details, health information and anything unrelated to choosing products.
Get consent to process personal data, kept clearly separate from other terms, and store it with the profile.
Show it and let shoppers delete it: a “What the assistant remembers about you” section in the customer account.
Ask rather than assume: “Shall I pick size 42 again?” instead of silently applying the filter.
Set a retention period and refresh outdated facts: a child’s size is different a year later.

Frequently asked questions

Does the model remember the shopper on its own?

No. A language model does not change its weights after a conversation with a shopper and remembers nothing between requests. “Memory” is whatever the application has stored itself and passed back to the model in the context: conversation history, a summary of preferences, past orders.

How is agent memory different from a CDP or an affinity profile?

A CDP and an affinity profile hold the store's behavioural data: views, clicks, purchases, interest in categories and brands. Agent memory adds what the shopper said in the conversation: “I'm looking for a gift for my sister, she's a size 38”, “I don't like wool”. It works best as one system: the profile feeds the assistant's memory, and facts from the conversation enrich the profile.

Can a store use what ChatGPT remembers about its customer?

No. ChatGPT's memory is stored by OpenAI and controlled by the user: it can be switched off, individual memories can be deleted, and temporary chats run without memory. The store sees only what reaches it in a specific request, so it builds its own memory on its own first-party data.

How does assistant memory fit with GDPR and CCPA?

Facts about a specific shopper are personal data, so processing them needs a lawful basis — for an assistant's memory, usually consent. Under GDPR the shopper also has the right to erasure (Article 17), and the CCPA gives California consumers similar rights to know and delete. In practice: remember only what helps with product selection, show shoppers what is stored and let them delete entries.