AI agent memory types: what an agent remembers and for how long
AI agent memory types come in two groups: what the agent keeps in mind during one conversation, and what is written down somewhere and looked up when needed. An AI agent is a software helper that reads, thinks and hands back a result. On its own it remembers nothing from one day to the next. Everything it seems to "remember" was written down for it.
Why this matters to a business
Picture a helpful temp who starts every morning with a blank mind. What they know about your company depends on three things: the note you leave on the desk, the folders they may open and the briefing they got on day one. An agent works the same way.
That is good news. Memory that is written down can be read, corrected and deleted. You are never left guessing what the agent has picked up.
The four memory types
| Type | What it holds | How long it lasts |
|---|---|---|
| The conversation | What you and the agent said in this chat | Until the chat ends |
| The summary | A short note that replaces a long chat | As long as the chat is kept |
| Documents and records | Your files, notes, customers and tasks | Until someone changes them |
| Instructions | The agent's job, tone and limits | Until someone edits them |
Short-term memory: the chat and its summary
Within one chat, the agent sees the earlier messages each time you write. A very long chat becomes slow and unfocused. Good systems therefore shorten it into a note. A useful note keeps your wishes, exact names and numbers, open questions and things still to do. It drops greetings and repetition.
A summary can lose detail. If the agent seems to have forgotten something from an hour ago, just give it the detail again.
Long-term memory: look it up
For anything that should last longer than one chat, the reliable way is simple. The fact sits in a document or a record, and the agent looks it up when a question needs it. This has three advantages.
- You can see the source. The answer points to a document you can open.
- You can correct it. Fix the document and the next answer is right.
- Access rules still apply. Someone who may not open the salary file does not get its content through the agent.
Where should a fact be kept? A sorting guide
- "Always write to customers in a formal tone" belongs in the instructions.
- "How our refunds work" belongs in a document the agent can search.
- "This order is due in March" belongs in the record of that order.
- "By the office I mean the Berlin office today" belongs in the chat and nowhere else.
The rule behind the list: if other people need the fact too, put it in a record or a document. A chat is not a filing cabinet.
Two questions to ask about any agent
- Does it keep people's chats apart? What one person told the agent should not show up in another person's answer.
- Where does the truth sit? It should sit in your own data. Memory is a convenience on top of it.
How Worknano handles this
Worknano's page assistants keep long chats useful with the Assistant Memory Summarizer. It shortens a conversation into a brief note that keeps preferences, exact record names, confirmed facts, open questions and pending actions. The next message starts from that note plus the latest messages. The summarizer adds no facts of its own. For long-term knowledge, assistants and agents use semantic search, a search that understands meaning and not only exact words. It covers documents, tasks, deals, contacts, OKRs and the text inside uploaded files, always within the permissions of the person asking. Your records stay the source of truth.
- Assistant Memory Summarizer: keeps long chats useful
- Semantic search: ask your workspace in your own words
Create a free workspace and see how an assistant keeps the thread in a long conversation.