Claude Memory Prompt: A Reusable Context Template

Claude does not remember previous conversations by default. Each session starts fresh. A memory prompt is a short block of text you paste at the start of a new conversation to give Claude the context it needs: the verified facts about the project, your working preferences, and where things currently stand. Without it, you spend the first minutes of every session re-explaining what you already covered.

Tools like fenbs solve this problem by storing that context outside any individual chat, so a connected assistant can read it automatically at the start of a session without you pasting anything. But if you are working in Claude directly without an MCP connection, a well-structured memory prompt is the practical alternative. This page gives you a template to copy and explains what makes a memory prompt effective.

What belongs in a Claude memory prompt

A good memory prompt contains three kinds of information. The first is verified facts: things that are true and do not change session to session, such as the project's goal, the technology stack, and the constraints that always apply. The second is working preferences: how you want Claude to respond, what format you prefer, what to check before suggesting a change. The third is current state: where the work stands right now, what was done last time, and what the next step is.

• Verified facts: project goal, stack, key decisions already made, hard constraints.
• Working preferences: response format, technical depth, review behavior, things to always or never do.
• Current state: what was completed last session, what is in progress, what the next action is.
• Open questions: anything that needs a decision before work can continue.

Keep the prompt compact. A very long prompt is slower to paste, harder to maintain, and may push important instructions past the point where Claude reads them most reliably. Aim for the information that would change the next action, not everything you know about the project.

The template

Copy this template, fill in the sections, and paste it at the start of a new Claude conversation. Update the Current state section at the end of each session before you close the chat.

## Project context
Project: [Project name and one-sentence goal]
Stack: [Key technologies, frameworks, tools]
Repo/location: [Path or URL if relevant]

## Hard constraints (always apply)
- [Constraint 1: e.g., "Never modify the database schema without a migration script"]
- [Constraint 2: e.g., "Use pnpm, not npm"]
- [Add others]

## Decisions already made
- [Decision 1: e.g., "We use Tailwind for styling, not CSS modules"]
- [Decision 2: e.g., "Auth is handled by the API; the front end never stores tokens"]
- [Add others]

## How I want you to work
- [Preference 1: e.g., "Ask before making changes to more than one file"]
- [Preference 2: e.g., "Keep explanations short unless I ask for detail"]
- [Preference 3: e.g., "Point out errors before starting a task if you see them"]

## Current state (updated [date])
Done last session:
- [What was completed]

In progress:
- [Task: where it stands]

Next step:
- [The specific next action to take]

Open questions:
- [Anything that needs a decision before work continues]

Verified facts versus working assumptions

The most common memory prompt mistake is including information that has not been confirmed. If you write "the API is deployed at [URL]" when you are not sure that is current, Claude will act on it confidently. Separate what you know for certain from what you believe to be true. A note like "I think the staging URL is [URL] but I have not checked today" is more honest and gives Claude a reason to verify before using the information.

Similarly, distinguish between decisions that are settled and preferences that you are still forming. An unsettled preference in the "Decisions already made" section may cause Claude to make a choice you later reverse. Mark open questions explicitly so Claude knows to ask rather than assume.

Keeping the current state section up to date

The template is only as useful as the "Current state" section is current. At the end of each session, update it before you close the chat: move completed items to done, describe where in-progress work stands, and write the next step precisely enough that Claude can start on it without asking.

You can ask Claude to draft the update for you: "Before we finish, summarize what you completed this session, what is still in progress, and what the next step should be." Review the draft and correct it, then save the updated prompt somewhere you can find it — a note-taking app, a text file in the project, or a skill in fenbs.

A persistent shared version

Pasting a memory prompt at the start of each session is a workaround for Claude's lack of persistent memory across conversations. A more durable setup is to store the project context in a tool that a connected assistant reads automatically. fenbs keeps projects, tasks, decisions, and Where we left off in one place. A Claude session with the fenbs MCP connection reads that context at the start without any manual pasting.

For a team, the shared version is especially valuable. Each person has their own conversations, but the project state in fenbs is common to all of them. See the Claude project memory guide for more on how Claude Projects can help with memory within a single workspace.

How to do this with fenbs

1. Create a free account. Start a project for the work you want to keep context on.
2. Add the verified facts and hard constraints as Decisions and rules. Add current tasks to the kanban board.
3. At the end of a session, ask your assistant to save Where we left off on the project and Where it stands on each task.
4. Connect Claude to fenbs from Integrations. The next session reads your rules, current tasks, and Where we left off automatically.
5. Keep the memory prompt template from this page as a backup for sessions where MCP is not available.

For related reading on preserving Claude context, see the Claude project memory guide and the Claude code memory guide. For a full picture of how shared memory works across a team, see Claude Code Projects shared memory.