AI agent memory is a set of saved notes about you, such as your role, your preferences and corrections you've given, that the agent loads into later conversations so you don't have to repeat yourself. It's a selection of what you've said, never a full recording, and it's only as useful as the notes are accurate.
Memory is the main thing that separates a chat tool from an agent, as AI usage vs an AI agent explains. It also means an agent can carry a wrong idea about you from one week into the next, which is why a little maintenance goes a long way.
What's the difference between memory and context?
Context is what the agent can see right now. Memory is what it saved earlier and can bring back later.
Context covers the current conversation, including anything you've pasted into it. It's temporary. When a conversation gets long, older parts may be dropped or summarised to make room, and Anthropic notes that accuracy and recall degrade as the context grows.
Memory is information stored about you from earlier conversations: your role, your preferences, decisions you made, corrections you gave. It carries over from one conversation to the next.
A handy picture: context is the desk in front of the agent, and memory is the filing cabinet behind it. If it forgets something from the start of a long thread, you've usually hit a context limit. If it forgets something you told it clearly last month, the note was probably never saved.
How does AI agent memory work?
The agent writes short notes to a store outside the conversation, then loads the relevant ones when a new conversation starts.
The details depend on the engine. In Hermes Agent, the engine Able° runs on by default, memory is two small files with a fixed character limit, and a note exists only if the agent saves it. The documentation warns that an agent saying "I've added that to my memory" is just text unless the save happened. New notes there take effect from the next session, not the current one.
What gets saved usually falls into these groups:
- Facts about you and your work: your role, business, team, regular commitments and the vocabulary you use. These stop it writing like a generic assistant.
- Preferences and style: answer length, whether you want options or a recommendation, the tone you use with clients, what you never want to see in a draft.
- Corrections and decisions: rules you set after a mistake and conclusions you reached together. This is the most valuable group, and the one people most often forget to maintain.
- Recent work: what you've been busy with, which helps continuity but turns into clutter if nobody tidies it.
What it doesn't remember is anything that never passed through it. Your files, other apps and conversations elsewhere are invisible unless you paste the relevant part in.
Once those notes build up, you explain your business and your tone once or twice instead of every time. You can come back after a fortnight, write "pick up the supplier issue", and carry on without re-briefing.
How does memory go wrong?
Usually through something that was true once. Maeve manages fit-outs for a joinery firm in Harrogate. In March she mentions that the Ripon shop job is "wrapping up", and her agent notes it. The job is then extended by six weeks, and she never thinks to say so. When she asks for a status summary in April, the agent writes as if the work has finished. No error message appears, and the summary looks perfectly plausible, which is exactly why she nearly sends it.
The other common slips work the same way. A request made during a hectic week, such as "keep everything brief", hardens into a permanent rule that also trims the proposals that needed detail. Or the agent fills a gap with a reasonable guess, saves the guess, and repeats it later as if you'd told it. In each case the memory reflects what was said or assumed at one moment, long after that moment has passed.
Habits that keep memory accurate
- Say it explicitly. Don't expect the agent to infer that you prefer bullets or that a project has closed. Tell it: "From now on, use bullets for summaries" or "The supplier project is finished. Don't bring it up again."
- Correct the rule as well as the draft. If a reply is too formal, fixing that one reply teaches nothing. Say "I never use 'kindly' with clients. Use plain phrasing instead", and it becomes a standing rule. How to correct an AI agent goes further.
- Ask what it knows. Every week or so, ask "What do you remember about how I work?" and read the answer carefully. It's also the day-five exercise in your first week.
- Retire old facts on purpose. When something changes, say so and ask the agent to forget the old version. If two notes contradict each other, settle it yourself; otherwise the agent may pick either one.
- Keep sensitive material out. If you wouldn't want a note read out to a colleague, don't ask the agent to store it. Paste it into a conversation when you need it instead.
Where is AI agent memory stored?
With whoever runs the agent, usually on a server, often alongside your conversation history. Able° keeps it on your agent's own private server, in Singapore by default.
Ask any provider:
- Where is it stored, and in which country?
- Can I see and edit it?
- Can I delete all of it if I leave?
- If I delete a conversation, does what was learned from it go too?
People tend to get caught out by the last one. OpenAI says that in ChatGPT, deleting a chat doesn't automatically delete a saved memory created from it, and Claude's help centre says the same about memories generated from deleted conversations. You have to delete the memory separately.
If the answers are vague, treat memory as data you've handed over rather than a setting you control. What happens to your data covers the wider picture.
When should you skip memory?
If you only use an agent for occasional one-off jobs, memory adds little and carries some risk. It also pays to wait if you haven't yet decided what you want the agent to be good at, because memory reinforces whatever setup it's given, muddled or clear.
Even when it works well, memory isn't understanding. An agent can know you run a workshop and still suggest something that only makes sense in an office, so you still read what it produces.
Common questions
Does an AI agent remember everything you tell it?
No. It keeps a selection of notes, and some engines deliberately cap how much they store. A detail mentioned once in passing may never be saved, and older notes can be trimmed or replaced. If something matters, ask the agent to remember it, then check later that it did.
Is AI agent memory the same as training the model?
No. Memory is a set of notes stored for your account and loaded into your conversations, and it doesn't change the underlying AI model. Whether a provider uses your conversations to improve future models is a separate question, governed by its own settings and policies, so check those as well.
Why does my AI agent forget things?
Usually for one of three reasons: the detail was in a long conversation and dropped out of view, it was never saved to memory, or a newer note replaced it. Repeat the key fact clearly, ask the agent to remember it, and then ask what it remembers to confirm the note is there.
Should I store client details in my agent's memory?
Only what the agent needs to do the work, such as a client's name, preferred format or deadlines. Anything sensitive, such as health information, personal finances or confidential terms, is better pasted into a conversation when needed. Notes about identifiable people count as personal data under GDPR and most local data-protection laws, so keep them to what you'd be comfortable explaining.