How do you give an AI agent memory that holds up over months?
Build memory as a data system with rules: store each fact with its source and date, keep each person's or customer's memory separate, summarize old history, and decide which fact wins when two disagree. Retrieval puts stored facts into the prompt, and most of the ongoing work is updating those facts when they change.
What breaks after months of use
Storing and retrieving memories works well in the first weeks. The problems arrive with volume and time:
- Facts go out of date: a customer moved, a price changed, a policy was replaced.
- The same person or account is stored several times under slightly different names.
- Two memories conflict, and the agent picks the wrong one.
- Old, low-value history crowds out what matters in retrieval results.
- One person's information appears in another person's conversation.
- Several agents share memory, and one agent's mistake spreads to the others.
How we build memory
Each person or customer gets a separate search index, so one customer's history cannot be retrieved into another customer's conversation. In the memory systems we build for you, each stored fact keeps where it came from and when, which lets the system prefer the newer fact and show its source when asked. Old conversation history is summarized in layers, so months of use stay searchable without resending thousands of messages.
Permissions apply to memory the same way they apply to documents: an agent working for one team sees only what that team may see. We have built user memory and permissions for a production agent product. Our own products carry months of history per user, with a separate search index for each person and older history summarized in layers.
Memory and context are different jobs
Memory is stored in a database, and context is the part of it the model receives on a given call. Most memory failures happen at the step between them, when retrieval brings back the wrong items, too many items, or an old version of a fact.
So we test retrieval directly, with eval cases where the right answer depends on a recent fact that replaced an older one, where two customers have similar names, and where the agent should say it does not know. Traces should record which memories were retrieved on each run, so when an agent gets something wrong, we can see whether the stored fact was wrong or the retrieval picked the wrong one.
Getting this set up
If you want this set up for an agent you already run or one you plan to build, start with the one-week audit. It ends with a ranked plan and something working by Friday, and larger builds are quoted in writing after it.
Questions
What is the difference between agent memory and RAG?
Retrieval-augmented generation (RAG) is a method for fetching stored text into a prompt. Memory is the stored information itself and the rules for keeping it accurate: what is saved, how it is updated, whose it is, and which fact wins in a conflict. Most memory systems use retrieval as one part.
How do you stop an AI agent from remembering wrong information?
Store each fact with its source and date, prefer newer facts over older ones, merge duplicate records, and test retrieval with cases where a fact has changed. Traces show which memory the agent used, so a bad fact can be found and corrected.
Should we use a memory framework or build our own?
A framework can be a good start. Plan to build the parts specific to your business yourself: how records are matched, which sources are trusted, who may see what, and how outdated facts are replaced.