Claude Memory vs Context: What Each One Keeps

Claude memory is selected information kept for use across conversations; context is the material available for the current response. A fact can be saved without being used correctly in a particular answer, and a document can be in the current conversation without becoming a lasting memory.

This distinction explains why “I told Claude yesterday” and “Claude has the source now” require different checks. Tools such as fenbs keep an explicit project record outside a private conversation so you can ask an assistant to read the relevant work on purpose.

What is different about remembering a fact and using it?

Think of a note in a filing cabinet and the papers on your desk. The note can survive for months, but it helps today’s decision only if you find the right one and interpret it correctly. The papers on your desk can be useful now even if you put none of them away afterward.

That analogy has limits: an AI app decides how to assemble material for a model, and its internal behavior is not visible in full. Still, it gives you two useful questions. Where is this fact kept, and what evidence did this answer actually use?

For example, a studio might have a remembered preference for concise client updates. Today’s task also needs the client’s approved launch date. The preference helps the writing style, but it cannot supply a date that was never saved or provided.

What does Claude save between conversations?

Anthropic’s memory help page↗ (opens in a new tab) describes remembered topics that you can inspect, edit, or delete in Settings, then Memory. Those entries are selected information, rather than a guarantee that the complete text of every previous conversation becomes available in every new chat.

Persistent information can also have a scope: a boundary around where it is used. Claude’s project and account memory settings matter when you move between kinds of work. A shared project document, a remembered fact, and an instruction are separate things to check, even if they describe the same client.

Do not reduce that to “everything in a project stays isolated.” Current account-memory controls affect the boundary. Our Claude project memory explanation covers the feature, and the official help page is the place to verify the controls available to your account.

Choose lasting entries for information you would still endorse in a later conversation. A writing preference may remain useful; a draft plan may become obsolete after one meeting. Persistence makes corrections important because yesterday’s reasonable assumption can become today’s confident error.

What goes into the context for the current task?

The context window is the model’s limited working space. Tokens, small pieces of text used for measurement, describe how much material it processes. Anthropic’s usage and length guide↗ (opens in a new tab) separates that conversation-length constraint from account usage allowances.

For the reader, the useful distinction is between a source being somewhere in the app and being available in a usable form for this request. Point Claude to the approved document and ask it to read the relevant portion. Merely naming an old chat does not prove the model has the evidence you need.

More material is not always better. Two versions of the same brief can create a conflict, even if both fit. Identify which version governs, keep its date or revision visible, and explain whether the older version is background or has been replaced.

• For a writing task, provide the audience, approved source, and format.
• For a planning task, provide the current goal, constraints, and decisions still open.
• For a check, provide the proposed answer and the source it must agree with.
• For continuing work, provide verified progress and the next action rather than every abandoned idea.

Why can Claude miss something it has already seen?

Several explanations fit that symptom. The fact may exist only in another conversation. It may be saved incorrectly. The current answer may rely on an older source, or Claude may fail to apply information that was available. A confident response does not tell you which explanation is true.

Test the source before guessing about the mechanism. Ask Claude to identify the document or entry supporting the disputed claim, and distinguish quoted evidence from its own inference. If it cannot point to the current approved source, supply that source and ask for a corrected answer.

Keep the test small. Ask for one current deadline, one rule, and one next step rather than a complete account of the project. Check the answer against your record. A vague “yes, I remember” is not evidence that the right details will guide the work.

When the disputed fact is a team decision, keep the approved wording with who decided it and why. fenbs puts Decisions and rules beside the task record, so a connected assistant has a specific record to consult rather than reconstructing agreement from old discussion.

Does Claude Code use memory and context the same way?

The underlying distinction still helps, but the storage mechanisms differ. Claude Code reads local instruction files and can write automatic memory notes. Anthropic’s Code memory documentation↗ (opens in a new tab) describes those mechanisms; they are separate from the chat app’s saved topics.

In Code, a file on disk is persistent storage. When its contents load into a session, they also become context. This is why “the file exists” and “Claude read the file” are separate claims. The Claude Code memory guide covers the file-based details.

Long Code conversations can also be compacted, meaning their history is summarized to make room. Anthropic’s context-window guide↗ (opens in a new tab) explains how different material reloads or is summarized. Do not assume a detail mentioned once in a long thread has the same status as an instruction deliberately kept for future sessions.

Which information needs an exact record?

Use an exact reviewed record when changing a word changes the next action. Approval, scope, acceptance criteria, and unresolved questions deserve clear wording. Memory can make everyday conversation smoother, while the work record explains what people actually agreed to do.

For a hypothetical studio, “the client likes simple pages” is background. “The client approved removing the optional phone field; message delivery still needs testing” is work state. The second statement identifies a decision and a remaining check without turning a style preference into permission for every design change.

Give the record a current home and link to supporting material. Avoid maintaining several equally authoritative summaries. If a decision changes, explain what supersedes it; if a test has not run, keep that uncertainty visible. In fenbs, the decision and task records give the next session a place to check those distinctions.

The broader memory versus context guide discusses this distinction across assistants. For the separate practical question of where Claude’s entries and files live, see Claude memory location.

How to do this with fenbs

1. Create a free account and add a project for one active piece of work.
2. Write the current task, its next action, and what has actually been checked. Keep source references in the task or project notes.
3. Record the decisions a person made in Decisions and rules. Keep unresolved choices open rather than letting an assistant turn a suggestion into agreement.
4. Connect a supported AI app and ask it to read the project context before answering. Check its summary of the rules and remaining work.
5. Before stopping, save progress and Where we left off. Start the next conversation by reading that record again.

fenbs is not the right fit if you only want an assistant to remember personal preferences, and it does not automatically import your Claude conversations. It is useful when the project needs a visible, editable record that another session or supported assistant can read.

Use a simple final test: can a fresh session explain the approved outcome, the evidence, and the next action? If it cannot, correct the record or its retrieval instructions. The AI memory hub has related guides when your next question concerns a different kind of persistence.