Project Memory
Understanding how Claude remembers things within a Project — and what it doesn't — is essential for building effective workflows. Memory in Claude Projects has three layers: the knowledge base, custom instructions, and conversation history.
What Is Stored
Knowledge base (permanent, manual)
Files you upload persist indefinitely until you delete them. This is the most reliable form of memory — it doesn't expire, isn't affected by context limits, and you have full control. Use it for facts, documents, and context you want Claude to always have access to.
Custom instructions (permanent, manual)
Your Project instructions persist indefinitely until you change them. Behavioural preferences, constraints, and context you specify here are always in effect — this is the most reliable way to maintain consistent behaviour across sessions.
Conversation history (persistent, subject to limits)
All conversations within a Project are saved and accessible in subsequent conversations. Claude can reference earlier discussions — decisions made, preferences expressed, prior outputs. Subject to context window limits (see below).
How Far Back Claude Can Reference
Conversation history within a Project is stored but not unlimited in what Claude can actively recall:
- Within a single conversation: Claude can reference everything that happened earlier in the same conversation, up to the context window limit (200K tokens for most paid plans).
- Across conversations in a Project: Claude has access to the conversation history list, but loading many long conversations simultaneously can exceed the context budget. For very active Projects with hundreds of conversations, Claude references the most recent and most relevant history.
- Important facts and decisions: If a conclusion from a previous conversation is important enough that you'll want Claude to remember it indefinitely, consider adding it to the knowledge base or custom instructions — don't rely solely on conversation history.
Knowledge Base vs Conversational Memory
These serve different purposes and should be used deliberately:
Use the knowledge base for:
- Facts that must remain accurate and stable
- Reference documents (specs, guides, code)
- Decisions or context you want permanently available
- Information that might get buried in a long conversation history
Rely on conversation history for:
- The natural flow of ongoing work
- Context built up through discussion
- Drafts and iterations in progress
- Short-to-medium term continuity (recent sessions)
When Context Gets Too Long
Every conversation has a context window — the amount of conversation history Claude can actively process at once. When a conversation grows very long:
- Claude may no longer be able to reference the very beginning of the conversation accurately
- For Claude.ai, the interface will warn you when you're approaching context limits
- In agentic and API contexts, context compaction (auto-summarising older parts of the conversation) can extend effective context further
Practical mitigation: for long multi-session projects, periodically ask Claude to "summarise the key decisions and context from our work so far in this Project." Save this summary to the knowledge base so it's available even as the conversation history grows long.
Clearing Memory and Starting Fresh
Sometimes you want to start a new conversation without prior conversation context influencing Claude's responses — while still keeping the Project setup:
- Start a new conversation within the Project: Each new conversation within a Project starts with just the custom instructions and knowledge base — it doesn't automatically load all prior conversation history unless you ask for it. This is your default "fresh start."
- Reference only what you need: "In our previous conversation we decided X — continue from there." This loads only what you explicitly reference.
- Archiving conversations: If a long conversation is cluttering the Project history, you can delete it from the Project to reduce noise. The knowledge base and instructions remain.
Checklist: Do You Understand This?
- Project memory has three layers: knowledge base (permanent files), custom instructions (permanent behaviour), and conversation history (persistent but subject to context limits)
- For important decisions and facts you want permanently available, add them to the knowledge base — don't rely on conversation history alone
- New conversations within a Project start fresh with instructions and knowledge base, but can reference prior history on request
- For very long Projects, periodically summarise key decisions and save to the knowledge base to avoid losing important context
- You can start a "fresh" conversation within the same Project to avoid prior context influence while retaining your setup