Context Snapshots: How to Manage Claude's Context Window Across Long Sessions

Claude just got memory. Free for everyone, from March 2026. It remembers your name, your preferences, your writing style, and your recurring projects. If you've spent the last two years re-explaining yourself at the start of every AI conversation, that friction is mostly gone now. But there's a problem memory doesn't fix. Most people don't notice it until they're 30 messages deep into a complex task and something starts going wrong. It's the context window. Memory and context windows are not the same thing Memory is Claude's long-term store. It holds what it knows about you across separate conversations: your job, your preferences, your ongoing projects. That persists indefinitely. A context window is different. It's the rolling scroll of everything said in the current session: your messages, Claude's replies, any documents you've attached, any tool outputs. Every conversation starts fresh, and there's a hard limit on how much it can hold. As a session gets longer, two things happen. First, the oldest parts of the context get deprioritised. Claude technically has them, but they carry less weight in its attention. You'll notice this when Claude contradicts a decision you made fifteen messages ago, or forgets a constraint you established early on. Second, you eventually approach hard limits where responses degrade or content gets cut off. Memory can't fix this. Memory only holds what was explicitly stored as persistent preference or context, not the messy working state of an in-progress session. The half-formed decisions. The reason you rejected option B. The constraint you mentioned once and assumed was understood. The pattern: context snapshots A context snapshot is a compact block capturing the current state of a session: what you're working on, what's been decided, what's still open, what needs to happen next. You generate it at a natural break point, start a fresh conversation, paste the block in, and pick up exactly where you left off. Clean context window, full situational awareness. The difference from just asking Claude to "summarise the conversation" is structure. A good snapshot is formatted for re-ingestion, not for reading. Dense and precise. A few hundred words that carry what a thousand words of conversation context would, without the noise. Here's what one looks like:

WORKING ON: Draft blog post about context snapshots. Decided to lead with the memory vs context distinction as the hook. Tone: direct, personal, no bullet lists in the body.

DECISIONS MADE:
- Opening with the Claude memory launch as context, not as the main topic
- Length target: 700 words
- Not mentioning the skill by name until the end

CONSTRAINTS: Voice must match existing posts. No em dashes. No AI writing patterns.

NEXT: Write the second half of the draft, starting from the "When to use it" section.

You can do this manually. Just ask Claude to produce a context snapshot of the current session and it'll give you something usable. Or you can build a skill that generates a consistent format every time, which is what I did. When it actually matters You don't need snapshots for short sessions. A 10-message conversation to draft an email or debug a formula doesn't need context management. You need them when responses start feeling less precise than they did early on. Claude repeats itself, walks back a decision without realising it, or stops honouring a constraint you established 20 messages ago. The other trigger is when you've loaded a lot of new information mid-session and earlier context is getting crowded out. The less obvious use case is forking. If you're mid-session and realise the problem has two branches worth exploring separately, take a snapshot at the fork and open two conversations from the same base. No re-explaining, no lost context. Both threads start fully informed. Memory handles the long game. Snapshots handle right now. Memory is for things that don't change: your preferences, your recurring projects, the context about you that's always relevant. "Always know this about me." Context snapshots are for in-session working state. "Right now, we're solving this, and here's where we are." Between the two, you're not starting every conversation from zero and you're not losing the thread when a session runs long. The context-snapshot skill I use in Cowork handles this automatically. It detects when a conversation has branched and generates separate blocks for each thread. Use the pattern before the session degrades rather than after.