Having a long conversation without losing context

Having a Long Conversation Without Losing Context

In early 2023, the content team at HubSpot began using Claude to help draft blog posts, email sequences, and social copy at scale. The workflow seemed simple: open a chat, describe what you need, get output. But within weeks, a pattern emerged that frustrated the team. Writers would spend 20 minutes establishing context — the brand voice, the target persona, the campaign angle — and then, the moment they started a fresh conversation the next morning, all of that was gone. Claude had no memory of yesterday's decisions. Every session started from zero. The team was rebuilding the same scaffolding over and over, burning time that should have gone into actual editing and strategy.

This wasn't a bug. It was a fundamental feature of how large language models work. Claude, like ChatGPT and Gemini, operates within a context window — a fixed block of text that represents everything the model can 'see' at once. Within a single conversation, Claude remembers everything you've said. The moment you close that tab and start a new chat, the slate wipes clean. The HubSpot team wasn't doing anything wrong; they just hadn't yet learned to work with this constraint deliberately. Once they did — building reusable context blocks, structuring long conversations intentionally, and knowing when to start fresh versus when to continue — their output quality jumped and their setup time dropped by roughly 60%.

The tension the HubSpot team hit is the central challenge of this lesson: Claude's within-conversation memory is powerful, but it has edges. Knowing where those edges are, and how to design your conversations around them, is what separates users who feel like Claude keeps forgetting them from users who feel like Claude is a fully briefed collaborator. That shift isn't about prompting tricks. It's about understanding a core mechanic and building a simple discipline around it.

The Context Window Is a Workspace, Not a Filing Cabinet

Think of a conversation with Claude as a shared whiteboard. When you start a session, the whiteboard is blank. Everything you type — your instructions, your background, your examples, your questions — gets written on that whiteboard. Claude can see all of it, every time it responds. This is why long conversations feel coherent: Claude isn't 'remembering' in the way humans do; it's re-reading the full whiteboard before every reply. The implication is significant. Early messages in a conversation carry real weight. If you established that your audience is 'senior procurement managers at Fortune 500 companies' in message two, Claude is still factoring that in at message forty-two — because it's all still on the whiteboard.

McKinsey's digital team ran into a vivid version of this when using Claude for competitive analysis work. A consultant would open a conversation, paste in a 3,000-word briefing document about a client's industry, then ask a series of increasingly specific questions over 90 minutes. The conversation worked beautifully — Claude's answers stayed tightly calibrated to the briefing. The problem came when a second analyst tried to 'continue' the work by starting a new conversation and asking follow-up questions without re-pasting the briefing. The responses were technically correct but generically framed, missing the nuance of the client's specific context. Same tool, wildly different output quality — because the whiteboard was empty.

The lesson the McKinsey team extracted: context isn't ambient. You don't 'have' context with Claude the way you have context with a colleague who attended the same meetings you did. Context is something you actively load into the conversation. The good news is that loading it is fast and the returns are immediate. A well-constructed opening message — one that establishes role, goal, constraints, and background — sets the entire conversation up for precision. Everything Claude produces after that opening is shaped by what you put in it.

What Happens When Context Is Thin

The editorial team at The Atlantic started experimenting with Claude for research summarization in late 2023. Journalists would paste article drafts and ask Claude to identify gaps in their argument or suggest additional angles to investigate. When it worked, it was remarkable — Claude would catch logical leaps, flag missing counterarguments, and surface adjacent story angles the journalist hadn't considered. But when it didn't work, the feedback felt frustratingly generic. 'Consider adding more data to support your claim' is not useful to a journalist who has already read 40 papers on the topic. The difference, the team realized, wasn't Claude's capability. It was whether the journalist had told Claude what they already knew.

An Atlantic staff writer who cracked this started every research conversation with what she called a 'known unknowns' header — a brief paragraph telling Claude exactly what she'd already researched, what sources she'd read, and what specific gap she was trying to fill. Instead of asking 'What angles am I missing on this story about urban heat islands?', she'd ask: 'I've reviewed the 2021 EPA urban heat study, three city-level case studies from Phoenix, Houston, and Atlanta, and a dozen peer-reviewed papers on cooling infrastructure. I have not yet looked at the insurance industry's angle or the intersection with urban planning zoning law. What am I likely missing in those two areas specifically?' The quality difference in Claude's responses was immediate and significant.

Comparing Conversation Approaches

The Analyst Who Never Repeated Himself

A financial analyst at Bain Capital Ventures developed a workflow that eliminated the re-briefing problem entirely. For every active deal he was tracking, he maintained a plain-text 'context block' in Notion — a 200–400 word document capturing the company name, sector, stage, key metrics, his current thesis, open questions, and the last three decisions he'd made in his analysis. Every time he opened a new Claude conversation related to that deal, he pasted the context block first. His first message was always the context block. His second message was the actual question. The analyst estimated this took 90 extra seconds per session and saved him 15–20 minutes of re-establishing the baseline and getting answers calibrated to his specific situation rather than generic VC analysis.

The context block approach works because it converts ephemeral conversation history into a portable, reusable asset. You're no longer dependent on keeping a single chat thread alive indefinitely — which creates its own problems, like threads that become unwieldy after 60+ exchanges. Instead, you extract the essential context from each session, update your block, and carry it forward. Claude treats a well-written context block exactly like it would treat a well-written opening message: as the authoritative frame for everything that follows. The analyst's Notion database of context blocks became, effectively, a project memory system that Claude could tap into whenever needed.

What This Means for How You Structure a Long Conversation

A long conversation with Claude isn't just a series of questions — it's a document that accumulates structure and meaning as it grows. The early messages are the foundation. If you establish your role, your goal, your constraints, and your style preferences in the first two or three exchanges, every subsequent response from Claude is built on that foundation. This is why experienced Claude users invest in their opening message even when they're in a hurry. A two-minute investment in a strong opening saves ten minutes of correction and re-prompting later. Think of it as writing a brief for a contractor — the clearer the brief, the less back-and-forth you need.

Long conversations also benefit from explicit checkpoints. When you've been working through a complex problem for 20+ exchanges, it's useful to occasionally ask Claude to summarize the decisions made and the current state of the work. This does two things: it confirms Claude's understanding matches yours, and it creates a natural summary you can paste into your context block when you're done. Some users do this every 15–20 messages as a habit. It's not that Claude loses track — the whiteboard is still full — but a summary forces any misalignment to surface before it compounds.

The third structural discipline is knowing when to start a new conversation rather than continuing an existing one. Very long threads — say, 50+ exchanges covering multiple distinct topics — can create a kind of context noise, where older, less relevant information from earlier in the thread competes with more recent instructions. If you've genuinely shifted to a new task or a new angle on a project, starting fresh with a new loaded opening is often cleaner than trying to redirect a thread that has accumulated a lot of prior context pointing in a different direction. The skill is recognizing that pivot point — and not defaulting to either extreme of always continuing or always starting over.

Key Principles So Far

1. Claude's memory is within-conversation only — every new chat starts blank, regardless of what you discussed yesterday.
2. The context window is a whiteboard, not a filing cabinet — Claude re-reads everything in the thread before each response, so early messages shape every answer that follows.
3. Context is not ambient — you actively load it, and the quality of what you load determines the calibration of every response you get.
4. A strong opening message is the highest-leverage investment in a conversation — two minutes of setup saves ten minutes of correction.
5. Context blocks solve the cross-session problem — a saved, reusable briefing document lets you reload full project context in seconds at the start of any new chat.
6. Periodic summaries within long conversations surface misalignment early and produce reusable artifacts for your context block.
7. Knowing when to start fresh versus continue is a skill — very long threads with multiple topic shifts can create context noise that a clean restart resolves.

What to Carry Forward

• Claude's 200,000-token context window is large, but what matters more than size is what you put into it from the start.
• The HubSpot, McKinsey, Atlantic, and Bain examples all point to the same root cause: output quality degrades when context is thin, and improves sharply when context is rich and explicit.
• The context block habit — build once, paste first, update after — is the single most practical system for maintaining quality across multiple Claude sessions on the same project.
• Treat the opening message of any important conversation like a contractor brief: specific, complete, and worth the time it takes to write properly.
• Explicit mid-conversation checkpoints ('summarize what we've decided') are both a quality check and a productivity tool.

When Context Collapse Kills Productivity

In 2023, the content team at HubSpot ran an internal experiment. They were using ChatGPT to help draft a 12-part email nurture sequence for a new product launch. The process started smoothly — the AI nailed the tone, understood the buyer persona, and produced a strong first email. But by email seven, something had gone wrong. The AI had forgotten the persona details established at the start. It started writing for a generic 'marketing manager' rather than the 'revenue-focused VP' they'd defined. The sequence became inconsistent. The team spent more time correcting drift than they saved by using AI. The root cause wasn't the AI's quality — it was that no one on the team understood how context actually works inside a long conversation.

This is the central tension in long AI conversations: the further you get from the beginning, the more the model is working from a compressed, fading picture of what you established early on. Claude doesn't 'remember' your opening instructions the way you do. It processes your entire conversation as a single block of text — called a context window — and as that window fills up, earlier content gets weighted less heavily in shaping the response. The HubSpot team's mistake was treating their conversation like a persistent workspace when it's actually more like a whiteboard that slowly gets erased from the top as you write at the bottom.

The Anchor Technique: Keeping Claude Oriented

A product team at Figma's growth division discovered a workaround through trial and error. They were using Claude to help develop a go-to-market strategy across multiple sessions. Their breakthrough came when a senior strategist started treating the opening message of every session like a project brief — dense, specific, and structured. She'd open with the company context, the specific goal for that session, constraints, tone, and any decisions already made in prior sessions. She called it 'the anchor.' The conversation that followed was dramatically more consistent because Claude always had a rich, current reference point at the top of the context — not buried in the middle of a previous exchange.

The anchor technique works because it exploits how Claude processes input. When your clearest, most complete framing sits at the start of the conversation, it shapes every response that follows. Think of it as setting a strong prior. If you've told Claude that your audience is skeptical CFOs who distrust vendor claims, every response it generates will be filtered through that lens — as long as that instruction remains accessible and hasn't been diluted by dozens of exchanges since. The Figma team started keeping a shared document of their 'anchor text' that they'd paste at the start of each new session, updating it as the project evolved.

This matters especially when multiple people are using Claude for the same project. Without an anchor, each team member starts a fresh conversation with their own framing, and Claude produces subtly different outputs — different tone, different assumptions, different level of formality. With a shared anchor pasted at the start of each session, the outputs converge. The Figma team reported that this single habit cut their editing and alignment time by roughly a third on that project. It's a simple technique, but it requires understanding why it works — which is exactly what most users skip.

A Different Industry, the Same Problem

Shift from tech to law. A mid-size UK employment law firm began using Claude to help associates draft client advice letters — a high-stakes context where consistency and accuracy matter enormously. The first associate to try it loved the results for short, self-contained tasks. But when she tried to use Claude to draft a multi-section advice letter covering redundancy, TUPE regulations, and settlement agreements in one conversation, she ran into trouble. By the time she reached the third section, Claude was producing advice that contradicted positions it had taken in the first section. The letter would have been unprofessional to send without significant revision.

The firm's solution was to treat each major section as a separate conversation with its own anchor — but they also introduced a summarization step between sections. After completing section one, the associate would ask Claude to produce a 'decisions summary': a bulleted list of every position taken, assumption made, and fact relied upon. She'd paste that summary as context at the start of the next section's conversation. This created a reliable thread between sessions without relying on Claude to maintain it automatically. The firm now uses this two-step method — complete a section, extract a summary, use the summary as the anchor for the next — as a standard workflow for complex documents.

Comparing Approaches to Long Conversation Management

How a Management Consultant Rebuilt Her Workflow

Priya Mehta, a senior consultant at a boutique strategy firm in Singapore, was skeptical of AI tools until a junior colleague showed her how to use Claude for client deliverable prep. Her first serious attempt was a 40-slide strategy deck for a retail client — a complex project with lots of moving parts. She started by dumping everything into one conversation: client background, market data, hypotheses, slide outlines. The conversation grew to over 60 exchanges. By the end, Claude was producing slides that felt disconnected from the strategic narrative she'd established at the start. The deck needed a full day of manual coherence editing before it was presentable.

Priya's second attempt used a completely different structure. She broke the deck into five logical phases: situation analysis, market sizing, strategic options, recommendation, and implementation roadmap. Each phase got its own conversation, opened with a tight anchor that included the client context, the overarching strategic recommendation (once that was locked in), and the specific output needed for that phase. She also ended each phase by asking Claude to write a one-paragraph 'phase summary' in the voice of the deck — essentially a transition paragraph she could paste into the next session's anchor. The result was a deck with a consistent narrative thread that needed only light editing. Her estimate: four hours saved on a project that had previously consumed a full working day of AI-assisted work.

What This Means When You're Working in Real Time

Most professionals use Claude reactively — they type a question, read the answer, type another question. This is fine for simple tasks. But for anything that spans more than ten exchanges or involves building something complex, reactive use is the wrong mode. The professionals who get the most out of long conversations treat the opening of each session as a design decision. They spend two or three minutes writing a proper anchor before sending a single task prompt. That upfront investment pays back in reduced correction time, more coherent outputs, and fewer frustrating moments where Claude seems to 'forget' what you told it.

The other real-time habit that separates skilled users from struggling ones is explicit confirmation. After Claude produces a significant output — a section of a document, a strategic framework, a set of decisions — skilled users ask Claude to confirm its understanding before moving on. A prompt like 'Before we continue, tell me the three constraints you're working within for this project' takes ten seconds to send and immediately reveals whether Claude has a clear picture of the guardrails. If the answer is vague or incomplete, you re-anchor before investing more time in the wrong direction. Think of it as a cheap quality check built into the workflow.

There's also a practical decision to make about when to start fresh versus when to continue. The general rule: if you're shifting to a meaningfully different task within the same project, start a new conversation with a fresh anchor. If you're continuing the same task with refinements, stay in the current conversation. The mistake most people make is staying in one conversation out of convenience — because the history is there and starting fresh feels like losing progress. But that conversation history is only useful to Claude if it's still in the active context window and hasn't been diluted. A fresh, well-anchored session almost always outperforms a long, drifting one.

Principles Extracted From These Cases

1. Claude processes your conversation as a single block of text, not as a persistent memory — earlier content gets underweighted as the conversation grows.
2. The anchor technique — opening every session with a dense, structured project brief — is the single highest-leverage habit for maintaining consistency in long conversations.
3. Summarization bridges the gap between sessions: asking Claude to produce a 'decisions summary' at the end of one phase gives you a ready-made anchor for the next.
4. Breaking complex projects into phase-specific conversations with shared anchors produces more coherent outputs than trying to do everything in one long thread.
5. The mid-conversation reset (restating key parameters when drift appears) rescues sessions without losing progress — use it before deciding to abandon a conversation.
6. Explicit confirmation prompts — asking Claude to restate its working constraints — function as cheap quality checks that catch drift early.
7. Starting fresh with a strong anchor nearly always outperforms continuing a long, drifting conversation out of convenience.

Key Takeaways From This Section

• Context windows are large but not infinite — and information placement within them affects how heavily Claude weights it.
• The 'lost in the middle' problem is real: instructions buried in the middle of a long conversation get less attention than those at the start or end.
• Treat the opening message of every session as a design decision, not a throwaway greeting.
• Shared anchor documents are a team-level solution — they standardize Claude's outputs across multiple users working on the same project.
• The summarize-and-restart method is the most reliable approach for high-stakes, multi-section professional documents.
• Knowing when to start fresh versus when to continue is a skill — and most users default to continuing too long out of misplaced convenience.

When the Thread Breaks: Managing Context in Extended Claude Conversations

In 2023, the product team at Duolingo began using Claude to help draft and iterate on their in-app copy — the tiny motivational nudges, error messages, and onboarding flows that millions of users see daily. Early on, a copywriter named Priya ran into a recurring problem: she'd spend the first twenty minutes of a conversation establishing Duolingo's voice — playful, encouraging, never condescending — and then, forty or fifty exchanges later, Claude's responses would start drifting. The owl mascot's personality would flatten. Formal phrasing would creep in. She'd have to re-explain the brief from scratch, losing the momentum she'd built. The conversation wasn't broken. But the context had quietly eroded.

Priya's frustration wasn't a Claude failure — it was a workflow design failure. Claude processes every conversation as a single, continuous window of text. When that window fills up with earlier exchanges, the most distant context loses influence. Think of it like a whiteboard that only holds so much: when you keep writing, the oldest notes get crowded to the edge. Priya was treating Claude like a colleague who'd remember everything they'd ever discussed. Instead, Claude is more like a brilliant collaborator who reads the entire conversation each time — but whose attention is naturally weighted toward what's recent.

The Anchor Message: A Technique Born from Frustration

Priya's solution was elegant. She started every long working session by sending what she called an 'anchor message' — a dense, structured briefing that captured the project's non-negotiables in one place. Voice guidelines. Audience. Constraints. The specific deliverable. Then, every fifteen or twenty messages, she'd send a 'context refresh' — a shorter message that restated the core brief, summarized decisions made so far, and flagged the next goal. This wasn't repetition for repetition's sake. It was deliberate context maintenance, treating the conversation like a document she was actively managing rather than a chat she was passively having.

The results were immediate. Claude's responses stayed on-brand across sessions that ran ninety messages deep. More importantly, Priya found that the discipline of writing context refreshes forced her to crystallize her own thinking. She couldn't write a vague refresh — Claude would produce vague output. The technique created a feedback loop where clearer prompting produced better work, which in turn made the next refresh easier to write. What started as a workaround became a core part of her creative process.

A Legal Team's Approach: Structured Summaries as Context Anchors

McKinsey's legal operations consultants face a different version of this challenge. When helping clients review long vendor contracts, they use Claude to analyze clauses, flag risks, and draft redlines across sessions that can span days. Losing context mid-review doesn't just slow things down — it risks missing a dependency between clause 4 and clause 22 that only makes sense if Claude 'remembers' the analysis from two hours ago. Their solution: a running summary document updated every session, pasted at the top of each new conversation. It's low-tech and it works.

The summary document follows a fixed template: parties and key dates, clauses reviewed and their risk ratings, open questions flagged for client, and decisions confirmed. When a new session starts, the consultant pastes this summary as the first message, then continues where they left off. Claude treats the summary as ground truth for the session. No context drift. No re-explaining that the indemnification clause was already marked acceptable. The conversation picks up cleanly, every time.

The Strategist Who Mapped Her Thinking Out Loud

Tomas, a brand strategist at a mid-sized agency in Amsterdam, uses Claude differently from either Priya or the McKinsey team. He's not managing a long document review or a content production pipeline — he's thinking. His conversations with Claude are messy, exploratory, full of pivots. He'll start by asking Claude to pressure-test a positioning statement, then veer into competitive analysis, then circle back to messaging. By message 30, the conversation has covered so much ground that neither he nor Claude has a clean thread to pull.

Tomas's fix is to periodically ask Claude to do the context work for him. Every twenty messages or so, he sends a single prompt: 'Summarize the key conclusions we've reached so far and the current open question.' Claude produces a tight synthesis. Tomas reviews it, edits it for accuracy, and pastes it back as a new anchor. He's essentially outsourcing the cognitive load of context maintenance to Claude itself — then verifying the output before trusting it. It's a smart loop: Claude helps you stay organized inside the very conversation where organization is hardest.

What Priya, the McKinsey team, and Tomas share is a shift in mindset: they stopped treating conversations as something that happens to them and started treating conversations as something they actively architect. The mechanics are simple — anchor messages, context refreshes, running summaries, mid-conversation synthesis requests. The underlying principle is the same across all three: Claude's output quality in message 50 is a direct function of how clearly you've maintained the context leading up to it. Good conversation hygiene isn't overhead. It's the work.

This also reframes how you should think about starting a new conversation versus continuing an old one. Many professionals reflexively continue in the same thread, assuming continuity is preserved. But if a previous conversation is long and you're returning to it the next day, the context that matters most — your goals, constraints, and prior decisions — is buried under dozens of older messages. Starting fresh with a well-crafted anchor message often produces better results than continuing a cluttered thread. The conversation history isn't an asset if you haven't actively maintained it.

Claude 3's 200,000-token context window means you rarely hit a hard cutoff. What you're managing isn't a technical limit — it's an attention and coherence challenge. Long conversations accumulate noise: early tangents, discarded directions, superseded decisions. A context refresh clears that noise and re-centers Claude on what actually matters now. Professionals who master this skill don't just get better outputs — they get faster outputs, because Claude spends less time reconciling contradictory signals buried in the conversation history.

1. Claude's attention is weighted toward recent messages — foundational context set only at the start of a long conversation loses influence over time.
2. An anchor message at the start of every session — restating goals, constraints, and tone — is the single highest-leverage context habit you can build.
3. Context refreshes every 15–25 messages keep long single-session conversations coherent without requiring you to start over.
4. A running summary document, pasted at the start of each new session, is the most reliable technique for multi-day or multi-session projects.
5. Asking Claude to summarize the conversation so far outsources the cognitive load of context maintenance — but always verify the summary before using it as an anchor.
6. Starting a new conversation with a strong anchor often outperforms continuing a long, cluttered thread from the previous day.
7. Context management isn't overhead — it's the discipline that separates professionals who get consistently excellent Claude output from those who get inconsistent results.

• Write an anchor message before every significant Claude session — goal, audience, constraints, tone in under 200 words.
• Send a context refresh every 20–25 messages in long conversations to prevent drift.
• Use a running context log document for any project spanning multiple sessions.
• Ask Claude to summarize mid-conversation when things get complex — then verify and edit the output.
• Choose between continuing a thread and starting fresh based on context quality, not habit.
• Treat conversation architecture as a skill: the structure you create directly determines the output quality you receive.