Managing context and memory

What you’ll be able to do

Spot when a conversation has grown too long or cluttered to stay reliable
Restart cleanly with a summary that keeps the decisions and drops the noise
Persist stable context in the right place so you stop re-explaining it

What you need to know

Context is the working space, and it is finite

Within a conversation, Claude works from what's in that conversation: your messages, its replies, and any files or Project knowledge in play. That working space, the context, is large but not unlimited, and very long or cluttered conversations bring two practical problems.

Drift. After dozens of turns, many drafts and a few changes of direction, early instructions and decisions are competing with a lot of later material. You may notice Claude going back to a format you abandoned, forgetting a constraint you set at the start, or mixing up details from different versions.
Limits. A conversation can eventually get too long to continue. Exact limits vary by model and plan and change over time, so don't plan around a specific number. Plan around the habit of keeping chats focused.

Separate conversations are separate by default. A new chat does not automatically know what happened in the last one unless something carries it over, such as memory, a Project, or you pasting in a summary.

Continue, summarize-and-restart, or persist

The exam objective names three moves. Choosing among them is the skill.

Continue when the chat is focused on one task, is still producing good results, and the earlier context is still relevant.
Summarize and restart when the chat is long, has changed direction several times, or Claude is drifting from earlier instructions. Ask Claude to write a concise handoff summary covering the goal, the decisions made, the constraints and the open questions. Check it, correct anything wrong, and start a fresh chat with it. You keep what matters and drop the clutter.
Restart fresh, without a summary, when you're switching to an unrelated task. Carrying over irrelevant context only adds noise.
Persist when the same context is needed again and again, such as a style guide, a team glossary, standing preferences or project background. Put it somewhere that survives across chats instead of re-pasting it.

Where to persist: Projects and memory

Project knowledge and Project instructions are the right home for context tied to a specific body of work: the client brief, the policy documents, the house style for this newsletter. Every chat in that Project can use them, and they stay put until you update them (Domain 5).

Memory, where it's available and enabled, lets Claude remember useful context across conversations, such as your role or your preferred format. You can view and edit what it remembers. Memory suits personal, durable preferences. It's not the place for a reference document, and it's not something to rely on blindly. Check what's stored now and then, and correct anything outdated or wrong.

Some things shouldn't persist at all. Sensitive personal data, confidential one-off details and anything your organization's policy restricts should not be parked in memory or in shared Project files just for convenience (Domain 6). Availability of memory and the related controls varies by plan and admin settings.

Writing a good handoff summary

A restart only helps if the summary is good. Ask for it in a structure you can check quickly:

Before

Summarize our chat.

After

Write a handoff summary for a fresh chat: the goal, decisions we've made and why, constraints I gave you, the current draft's structure, and open questions. Keep it under a page and leave out options we rejected.

Then read it. A summary can drop or distort a decision, and if you paste it in unchecked the error carries forward with an air of authority.

Key concept

When a chat gets long or starts drifting, restart with a checked summary. When the same context is needed again and again, persist it in a Project or memory instead of re-pasting it.

Exam questions describe symptoms such as forgotten constraints, a reverted format, re-pasting the same files, or a new chat that "doesn't know" something. The right answer picks the move that fits the symptom: restart with a summary, persist the context, or simply start fresh.

Practice scenario

ScenarioA project manager has spent three weeks in a single chat planning an office move. The chat covers vendor quotes, a seating plan, several rejected timelines and a communications plan. Recently Claude has started proposing a move date the team ruled out two weeks ago and formatting updates in a style she asked it to stop using. She plans to keep a running list of corrections in the same chat.

Work it through, then open this

Adding corrections to an overloaded chat treats the symptom, not the cause. The long, multi-direction history is making earlier decisions harder to keep track of. A better move is to ask Claude for a structured handoff summary (goal, confirmed decisions including the ruled-out date, the preferred update format, and open items), check it carefully, and start a fresh chat with it. Because the move is an ongoing body of work with recurring reference material, she should also set up a Project holding the vendor quotes and the format guidance, so future chats start from the same stable context.

Hands-on exercise — Restart a long chat without losing anything

Beginner · 15 min

What you’ll learn

Writing and checking a handoff summary
Seeing the difference a clean context makes
Deciding what to persist and where

Pick a long chat you’ve had with Claude, or create one by working through several drafts and a change of direction, then ask for a structured handoff summary using the “After” prompt above.
- Why: A structured request makes missing or distorted decisions easy to spot.
- You should see: A short summary with labeled sections. Check it against the chat and fix at least one detail if anything is off.
Start a new chat, paste the checked summary, and ask Claude to continue the work.
- Why: This shows how a clean start with the right context compares with an overloaded thread.
- You should see: Claude working within the confirmed decisions without bringing back rejected options.
List what from this work you’d need again, and sort each item into “Project knowledge”, “memory” (if available) or “don’t persist”.
- Why: Persisting the right things in the right place stops the cycle of re-explaining.
- You should see: Reference documents going to the Project, personal preferences going to memory, and anything sensitive or one-off staying out of both.

Exam traps

Continuing one endless chat for weeks of unrelated work

Long, mixed conversations make earlier instructions harder to keep straight. Keep chats focused and start new ones for new tasks.

Starting a fresh chat without carrying over the decisions already made

A restart without a summary throws away the decisions you made along with the clutter. Carry over a checked handoff summary.

Assuming Claude remembers a previous chat when memory isn’t on and nothing was persisted

Conversations are separate by default. Without memory, a Project or a pasted summary, the new chat starts with no knowledge of the old one.

Re-pasting the same reference material into every chat instead of persisting it

Recurring context belongs in Project knowledge or instructions, where every chat in the Project can use it.

Repeating the instruction louder when Claude drifts in a long chat instead of resetting

When drift comes from an overloaded conversation, a clean restart with a good summary fixes the cause.

Letting sensitive details persist in memory or Project files without thinking about who sees them

Persisting is convenient, but sensitive or restricted information shouldn’t be stored just to save typing. Follow your organization’s policy.

Sources

Claude Help Center
Claude Docs
Anthropic, Claude Certified Associate – Foundations Exam Guide v1.0 (July 2026)