AI Memory in one minute
Think of the context window as the model's current worktable and memory as material kept elsewhere for possible later use. An application may bring selected notes back onto the table. Saving a note does not mean the model sees it in every interaction, and saying 'I will remember' does not prove that anything was stored.
AI Memory covers several different mechanisms
| Type | What might be retained? | Important qualification |
|---|---|---|
| Session state | Progress, earlier messages or intermediate results. | Visible history is not necessarily all included in the next model request. |
| Preferences | Preferred language, tone or output format. | Preferences need a scope and a way to correct them. |
| Persistent memory | Notes or task state kept between sessions. | Storage, retrieval and deletion depend on the application. |
| External data | Documents or records in connected systems. | Availability does not imply automatic inclusion in context. |
These categories can overlap. A preference may be stored persistently, while a task record may live in a database used by several sessions. They are a way to discuss behaviour, not a claim that every assistant provides four standard memory products. Some applications retain little beyond the current session, while others offer user-managed or organisation-managed stores.
The Context Window is the information the model can process within an interaction, subject to its limits. It can include instructions, conversation history, retrieved documents and tool results. A long chat displayed on screen may be summarised or partly omitted before the next request. The visible interface and the model's active input are different things.
Training adjusts model weights through a learning process. Saving a customer preference in an application database does not, by itself, retrain those weights. Likewise, attaching a document to a conversation supplies context rather than automatically becoming permanent model knowledge. These distinctions help explain why a saved fact may be available in one product or workspace and absent in another.
Saving, selecting and applying information
- Retain
Store a relevant preference, task status or source-linked note.
- Select
Retrieve the information appropriate to this user and assignment.
- Apply and check
Use it as context while checking whether it is still valid.
The storage can be a file, a database record or another controlled system. A vector index can help retrieve semantically related notes, but it is not required for every memory design. Exact task identifiers or structured fields may be more reliable for tracking which products have been reviewed. The useful design depends on what needs to be remembered and why.
Anthropic's memory-tool documentation provides one concrete example: client-side software performs storage operations requested through a tool. The application manages the actual files and their scope. This illustrates that memory can be implemented outside the model; it does not establish that all assistants use that mechanism or share the same persistence rules.
Consider AI Agents reviewing 400 product records. Useful retained state includes the task identifier, records already inspected, unresolved conflicts and the versions used for proposed changes. On resuming, the system can avoid repeating completed work. It should still check whether a colleague has changed relevant records since the earlier review.
Status labels matter. 'Draft prepared', 'approved' and 'written to the shop' are different states. A summary that collapses them into 'done' can cause a later step to publish unapproved content or skip unfinished work. Preserve the distinction between a model proposal, a human decision and a verified external action.
RAG and memory can overlap in how stored material is retrieved, but they describe different concerns. RAG emphasises retrieval as grounding for generation. Memory emphasises retaining information for later interactions or task continuation. An application may retrieve both company documents and earlier task notes; they should retain separate provenance and authority.
A remembered price is a good example of information that may need rechecking. It can explain what was seen during yesterday's analysis, but a current stock or price question should consult the designated live source through Connectors where available. Persistence is useful for continuity, not a reason to bypass freshness checks.
Scope, correction and safe reuse
Memory should be associated with the right person, organisation, customer and task. A preference from one client may be inappropriate for another. Confidential project notes must not become shared context simply because they are semantically similar to a later question. Retrieval needs access control as well as relevance ranking.
Make important memories inspectable and correctable. A useful record can include its source, date, scope and status. When a person corrects a preference or fact, the application should know which stored item is superseded. Keeping contradictory notes without a resolution rule can make future answers less reliable.
Deletion also has a scope. Removing a preference from the active store may not erase every backup, log or copy in an external system. Products differ in their retention controls. Explain what a deletion operation affects instead of promising that all traces disappear everywhere. The same care applies when an account is disconnected but earlier information remains in an index.
Stored material can contain misleading instructions. If an external document persuades the application to save an unsafe instruction as a preference, that instruction may influence later sessions. Treat source text as data, preserve its origin and prevent it from granting new permissions. Persistent storage should not turn untrusted content into an authoritative user request.
- Keep the minimum information needed for continuity.
- Separate source facts, drafts, approvals and completed actions.
- Associate records with the correct identity and task.
- Provide correction and deletion controls appropriate to the product.
- Recheck volatile information before consequential use.
An Agentic Workflow can preserve progress in durable checkpoints, while Human-in-the-loop resolves uncertainty before important actions. Neither replaces a clear record of what happened. Good AI Memory helps the system continue the right task with the right evidence, while remaining able to admit that an old note is incomplete or no longer current.