MMOOZZGG.com

Glossary

What is AI agent memory?

In short

AI agent memory is everything an AI agent remembers besides the current request: the course of a conversation, facts about a client, past decisions and lessons from mistakes.

AI agents · agent memory

MMOOZZGG.com editorsChecked: 1 min read

For a child

When you solve a long problem, you write a draft so you do not forget the answers along the way. In your diary you keep what you will need next week: the timetable, homework, which friend likes which games. A helper built on a neural network also has a draft and a diary. Without them, it would meet you for the first time every time.

In plain words

A language model remembers nothing between conversations by itself: it sees each request from a blank sheet. Memory is built around it. The short-term part sits right in the request: the latest messages and intermediate results. The long-term part is stored apart, in a database, and is added to the request when needed. For example, a client wrote to support in spring and said they pay only by bank transfer. Two months later the agent finds this note and offers the right payment method at once, without asking again.

For business

Memory makes an agent useful in long client relationships and in multi-step tasks: it does not ask the same thing twice, it keeps agreements in mind and it continues work after a break. But everything the agent remembers about people becomes personal data. Its storage falls under 152-FZ, Russia’s personal data law, so you need retention periods and a way to delete a record on request. The second cost item is volume: the more of the past goes into each request, the more each answer costs. The third risk is outdated or wrong records that the agent takes as truth. Before launch, decide what to remember, where the database lives and who may view it. These rules are set together with the setup of an agent that books clients.

For a specialist

Agent memory is split into working and long-term memory. Working memory lives in the context window: the dialogue history, the plan and the results of tool calls. Long-term memory sits in an external store, often a vector database or an ordinary table, and is retrieved by search, as in RAG. By content, there is episodic memory (events and past dialogues), semantic memory (facts about the user and the world) and procedural memory (skills and instructions). In MemGPT (Packer et al., 2023), the model itself moves data between memory tiers, much like virtual memory in an operating system. In Generative Agents (Park et al., 2023), agents keep a full log of experience in natural language, sum it up in reflections and use it for planning. Open questions remain: what to forget, how to resolve conflicts between records and how to keep a harmful record out of the store.

Example

A booking agent at a car service center remembers that a client drives a diesel car and last time asked for calls only in the afternoon. On a new booking, the agent offers suitable time slots and the right mechanic at once.

Related terms

Questions and answers

How is agent memory different from the context window? The context window is the amount of text a model sees in one request. Agent memory is wider: it includes an external store, and only the records the task needs go from it into the window.
Can you delete what an agent remembers? Yes, if the memory sits in your own database: you delete a record like any other. Plan for this in advance, because under 152-FZ, Russia’s personal data law, a person may demand that their data be deleted.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us