How to Give Claude Persistent Memory for Ongoing Work
Use native memory for useful recall, project knowledge for reference material, and an editable work record for decisions and progress. Test what a fresh chat can actually use.
Updated 5 min read
To give Claude persistent memory, enable its memory feature, review what it saves, and put exact project material in a place future chats can access. Persistent means the information survives a conversation; it does not mean Claude remembers everything or always retrieves the right detail.
For work that must continue across sessions or assistants, fenbs holds an explicit project record outside Claude. Use native recall for helpful background and a reviewed record for approved decisions, tasks and the next step.
How do you set up memory that survives a chat?
- Check Settings, then Memory, and enable memory if your account and organization allow it. Use an ordinary chat with memory available for the setup.
- Choose one stable, nonsensitive fact and tell Claude to remember it. Inspect the saved item under Topics instead of relying only on its reply.
- Create or open the project for your ongoing work. Add the reference files and instructions that should apply to its conversations.
- Write a short current-work record: what was approved, what is finished, what remains uncertain and what to do next.
- Start a new conversation in the intended project. Ask Claude to explain the relevant fact and the next action without supplying the answer again.
Anthropic’s memory guide (opens in a new tab) explains the settings and editable topics. Inspect your current settings rather than following an older screenshot or assuming that the feature is enabled.
Start with this small test before bringing over years of notes. You want to know what you saved, where it lives and whether a new session can use it. More material cannot compensate for a setup you have never checked.
Which information belongs in each place?
| Information | Practical home | What to verify |
|---|---|---|
| A general response preference | A native saved memory or appropriate instruction | The wording is correct and applies broadly. |
| An approved reference document | Project knowledge | The file is current and the chat is in the right project. |
| A recurring project instruction | Project instructions | Its scope is clear and competing instructions are removed. |
| A decision and unfinished next step | An editable work record | A person approved the decision and the progress is current. |
Say you are drafting a product handbook. “Use concise explanations” is a preference. The approved terminology file is reference material. “Legal must approve changes to the warranty text” is a rule. “The warranty section still needs review” is progress. Keeping those distinct makes corrections much easier.
A project’s title alone is not a working brief. Add the information Claude needs to the relevant knowledge or instructions. Anthropic’s project setup guide (opens in a new tab) shows those controls and explains that uploaded project knowledge is used across its chats.
How do you make the record useful instead of huge?
Write the smallest record that changes the next action. Include the goal, the authoritative source, the current decision and the next incomplete task. Keep old discussion as background rather than mixing every tentative idea into the instructions.
- State decisions as decisions: who chose, what they chose and what it applies to.
- Mark proposals and unanswered questions clearly so they do not become accidental rules.
- Record the result of a check precisely, including what has not been checked.
- Replace stale progress when the work moves forward; preserve the reasoning behind decisions separately.
For the handbook, “Draft complete; warranty review pending; do not publish” tells the next session more than a long celebratory summary. A future assistant should be able to distinguish a finished draft from permission to release it. In fenbs, Decisions and rules and Where we left off make those two jobs visible in the project context.
How do you prove it persists?
Test retrieval and use separately. First ask a fresh conversation what the relevant saved fact or project instruction is. Then ask it to do a small task that depends on that fact. The first checks access; the second checks whether the assistant applies it appropriately.
Keep the test neutral. “What still needs approval?” is stronger than “Do you remember that legal must review the warranty?” Check its answer against the record you can inspect. If it guesses, ask it to identify the source or say that it cannot find one.
Repeat this check after a meaningful correction or a change of account, workspace or project. You do not need an elaborate test suite; one realistic question and one small task can expose a mistaken assumption before it spreads into more work.
Make ongoing work a record you can inspect
Save the handbook’s approval rule, current tasks and remaining review in fenbs. A connected assistant reads the project context before starting and saves progress before stopping. That gives the next session an explicit starting point instead of depending on a remembered conversation.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
Does this also work in Claude Code or another assistant?
Claude Code has a different persistence system. Anthropic’s Claude Code memory documentation (opens in a new tab) describes instruction files and auto memory; changing chat memory is not the same setup. Follow Claude Code memory when the work happens in a code repository.
The wider Claude project memory guide covers project behavior, and Claude memory location owns the question of where saved information can be inspected. A one-time move from another app is covered in Claude import memory; it does not create continuing synchronization.
If several assistants need the same changing work state, keep one shared record that they deliberately read. Copying a summary into each app produces several versions to maintain. A new decision made in one conversation cannot reach the others unless someone updates what they consult.
How to do this with fenbs
- Create a free account and add one active project on your kanban board.
- Add its current tasks, concise AI context notes and the decisions a person has approved. Record continuing instructions as rules with a clear scope.
- Connect Claude and any other supported assistants using the setup guides in Integrations. The connection lets them use the outside record through MCP, a standard way for AI apps to use tools.
- Ask the assistant to read the project context before working and save progress before stopping. Test a fresh session by asking what remains and what needs a person’s decision.
fenbs is not the right fit for storing personal biographical memories or making Claude recall every sentence you have ever written. It keeps the work you choose to record, and it does not automatically import Claude memories or read your chat history.
Use the AI memory topic hub to explore the surrounding options. Your first useful milestone is modest: a new conversation accurately identifies the current decision and next unfinished action without you retelling the whole project.