Claude Memory Bank: Files, Skills, and MCP Options

A Claude memory bank is a deliberate record of the facts, decisions, and progress you want Claude to read again. You can organize it as files or a connected project record; the useful part is a reliable read-and-update routine, rather than the name of the folder.

Tools like fenbs keep that record outside the conversation, with project context and checkpoints available to supported connected assistants. This is different from hoping automatic memory selects every detail needed to resume your work.

How do you start with a small memory bank?

1. Choose one project and write its goal in a few sentences. State what is outside the current scope.
2. Record the decisions people have approved, including the reason and the person who chose each one.
3. List current tasks, separating completed work from work that still needs checking.
4. Write the exact next step, anything waiting on a person, and links to the source material.
5. Start a new session, have Claude read the record, and check its proposed next action before it works.

Do that before designing a large folder structure. A short accurate record beats many pages of unsorted chat summaries. If the next session can explain the goal, the applicable decision, and the unfinished check, you have the beginnings of a useful bank.

Use fictional or nonsensitive material for the first trial. For example, plan a community workshop rather than starting with a confidential client engagement. That lets you check the routine without making access or deletion mistakes expensive.

Should you choose files, a skill, or an outside tool?

These approaches can work together. A procedure can tell Claude how to read a record stored elsewhere. Choose a clear home for the current facts so a copied file and a connected tool do not become competing versions of the same decision.

Anthropic documents Claude Code instruction files and automatic memory↗ (opens in a new tab) separately. A folder you create yourself is not automatically the same thing as those built-in mechanisms. Tell the assistant where to look and verify that it actually reads the intended record.

What should the record contain?

Organize information by how the next session uses it. Put durable project facts in one place, decisions in another, and the current checkpoint where it is easy to find. The following is a suggested layout for files, not a required Claude format or a set of automatically recognized filenames.

project-record/
  start-here.md       Goal, scope, and links to current records
  decisions.md        Approved choices, people, reasons, and scope
  current-work.md     Tasks, status, evidence, and open questions
  checkpoint.md       Where to resume and what remains unchecked
  references.md       Source documents and why they matter

Write entries that preserve the difference between evidence and confidence. “Draft revised; the organizer has not approved it” is useful. “Draft finished” hides the remaining decision. Include a source reference when a statement matters enough that someone might need to challenge it.

Leave raw conversation transcripts out of the starting record unless they are necessary evidence. Link to the original material when you need it. In fenbs, projects, tasks, Decisions and rules, and Where we left off give these different facts explicit homes; how it works shows that arrangement.

What role can a skill or MCP connection play?

A skill can describe the method for maintaining the bank: read the starting record, check relevant decisions, work on the agreed task, then update the checkpoint. Anthropic’s skills explanation↗ (opens in a new tab) distinguishes procedural instructions from connections to outside data.

MCP, a standard way for AI apps to use outside tools, can provide access to a shared record instead of a local folder. The connection makes reading and updating possible; your working instructions still need to say when to read, what to save, and which changes require a person’s approval.

For developers, Anthropic also offers a client-side memory tool↗ (opens in a new tab). Claude requests operations while your application supplies the storage. That is an implementation route for a product, rather than an automatic connection to your Claude app history.

If your question is how procedures differ from remembered facts, use the separate memory versus skills guide. Keep the bank focused on current project truth rather than making it a second instruction manual.

How do you keep the bank accurate between sessions?

Make upkeep part of finishing the work. Ask the assistant to identify which tasks changed, what was checked, what failed, and the next action. Review that update while the evidence is still available. Do not wait until another session discovers that the last checkpoint was too vague.

• Separate an accepted decision from a suggestion that still needs approval.
• Replace an obsolete current-state statement and preserve decision history where needed.
• Keep a failed check visible, including where it stopped and what remains uncertain.
• Link to the current source rather than copying several versions of the same document.
• Remove unrelated personal details and secrets from the shared work record.

Say the workshop venue changes. Update the accepted venue decision and any tasks it affects. The checkpoint should explain the consequence, such as “directions need revision,” rather than leaving the next assistant to infer it from an old conversation.

How to do this with fenbs

1. Create a free account, free for one person with no credit card, and add the project you want to continue across sessions.
2. Add current tasks and record the approved choices in Decisions and rules. Use context notes for supporting project facts.
3. Follow the relevant setup guide on Integrations to connect Claude through browser sign-in.
4. Ask the connected assistant to read the project context before starting, then verify its summary and proposed next step.
5. Before stopping, save progress on the tasks and project, including what remains and what is waiting on a person.

fenbs is not the right fit for a private folder you only want to keep locally or for building a storage engine inside your own application. It also does not import Claude chat history automatically. Choose it when the work needs a shared record that supported assistants can read.

How can you tell whether the setup is working?

Run a fresh-session test after a meaningful change. Ask Claude to name the approved goal, one rule that constrains the work, the next action, and one thing that remains unverified. Compare those answers with the current record before authorizing more work.

If it fails, inspect the record and the reading routine separately. A clear fact in an unread file is an access or starting-instruction problem. An ambiguous checkpoint needs better writing. More stored text does not necessarily fix either one.

Keep built-in features distinct with the guides to Claude project memory and Claude Code memory. The AI memory hub covers the broader choices; the practical test for your bank is whether another session can continue correctly.