You gave ChatGPT everything it needed. Your company background, the project constraints, the stakeholder preferences, your preferred tone. The responses were sharp and useful. Then you started a new conversation, and it was like talking to a stranger.
This is the single most common frustration with ChatGPT, and it happens because of how context windows and AI memory actually work. The good news: there are concrete strategies to fix it, ranging from two-minute adjustments to a structural solution that eliminates the problem entirely.
Here are five strategies, ordered from simplest to most effective.
Why ChatGPT loses context in the first place
Before the fixes, it helps to understand why this happens. ChatGPT operates within a context window: a fixed amount of text it can "see" during a conversation. For GPT-4o, that window is 128K tokens (roughly 100,000 words). GPT-5 pushes this to 400K tokens. Sounds like a lot, and it is, for a single session. But when you close that conversation and open a new one, the context window resets to zero. Every token from the previous session is gone.
ChatGPT's memory feature, introduced in 2024, is supposed to bridge this gap. But it has a hard ceiling of roughly 1,500 to 1,750 words (increased from around 1,200 words after the February 2025 update). That's about three pages of text. When you've told the AI hundreds of facts about your work across dozens of conversations, compressing all of it into three pages means most of the nuance disappears.
And memory isn't even reliable. In February 2025, a bug wiped memories for a significant number of users. Months of carefully built context vanished overnight with no recovery option. In November 2025, another memory incident occurred, though OpenAI acknowledged and resolved it within about 24 hours.
So you're working with a system that forgets everything between sessions, retains a tiny fraction of what you've told it, and occasionally loses even that.
Here's how to work around it.
Strategy 1: Write better custom instructions
Custom instructions are text that ChatGPT prepends to every conversation. You get roughly 1,500 characters per input box, totaling about 4,500 characters across all sections. This is the lowest-effort, highest-impact change most people can make.
The problem is that most people either leave custom instructions blank or fill them with vague statements like "I'm a product manager." That wastes the most valuable real estate in your ChatGPT setup.
Instead, pack your custom instructions with the specific context that you find yourself repeating:
- Your role and company. Not just "product manager" but "Senior PM at a Series B healthcare startup, 14-person team, B2B SaaS targeting rural clinic administrators."
- Your preferences. Writing tone, formatting standards, terminology choices. If you always have to say "don't use bullet points for everything" or "use metric units," put it here.
- Common tasks. "I frequently ask for product spec reviews, competitive analysis, and meeting agenda drafts."
- Key terminology. Terms specific to your industry or company that the AI might misinterpret.
Review and update your custom instructions monthly. As your projects and priorities shift, so should your baseline context.
Custom instructions are the minimum viable context strategy. They're not enough on their own, but everything else builds on top of them.
Strategy 2: Use ChatGPT Projects
Projects are ChatGPT's most underused feature. They let you upload reference files and write project-specific instructions that persist across every conversation within the project.
The file limits depend on your plan: 5 files for Free, 25 for Plus, and 40 for Pro, with a 512MB cap per file. This means you can upload your product roadmap, style guide, technical architecture document, and company overview, and ChatGPT will reference them in every conversation you have within that project.
The key to making Projects work well:
Keep conversations focused. After about 30 messages, start a new conversation within the same project. Write a brief handoff summary at the end of each conversation: "We decided X, Y, and Z. Next conversation should focus on implementing Z." This gives the next conversation a clean starting point while your project files provide the persistent context.
Organize by workstream, not by topic. Create projects around your actual workflows: "Q1 Product Launch," "Engineering Architecture Review," "Customer Research." Each project gets the specific files and instructions relevant to that workstream.
Update files when they change. Project files are static uploads. When your roadmap changes or your team structure shifts, replace the outdated files. Stale context is worse than no context because the AI will confidently reference outdated information.
Projects are significantly more effective than custom instructions alone because they give the AI access to full documents rather than a compressed summary. But they're still siloed per project, and you lose conversation-level context when you start a new chat.
Strategy 3: Structure your conversations deliberately
Most people interact with ChatGPT as a stream of consciousness. They type whatever comes to mind, assume the AI is following along, and then get frustrated when it loses the thread.
A better approach: treat every conversation like a briefing document.
Start with a context block. At the top of every new conversation, write a 3-5 sentence summary of what you're working on and what you need. Even if you're in a Project with uploaded files, an explicit context block at the start helps the AI prioritize the right information.
Use explicit references. Instead of "the thing we discussed," write "the pricing change from last week's product review (switching from per-seat to usage-based billing)." The AI has no memory of "last week." Every reference needs to be self-contained.
Summarize decisions periodically. Every 10-15 messages, write a brief summary of what you've decided so far. "To recap: we're going with Option B for the API redesign, targeting a March launch, and deprioritizing the mobile SDK until Q3." This pushes the key information toward the end of the context window, where the AI pays the most attention.
This last point matters because of a well-documented phenomenon called "lost in the middle." Research from Stanford and Berkeley (2023), confirmed repeatedly through 2025, shows that LLMs recall information best at the beginning and end of the context window. Information in the middle is the most likely to be missed. Periodic summaries exploit this pattern by refreshing your key decisions at points where the AI is paying attention.
Break long tasks into multiple conversations. If a task will take more than 20-30 messages, plan your handoff points in advance. At each break, write a summary that the next conversation can use as its starting context.
Strategy 4: Use Saved Memories strategically
ChatGPT's Saved Memories feature stores roughly 1,500 to 1,750 words of information that persists across all conversations. After the February 2025 expansion, this is enough to be genuinely useful, if you're deliberate about what goes in.
Do not waste memory on trivia. ChatGPT will try to save things like your pet's name or your favorite restaurant. That's a poor use of the most limited persistence layer you have.
Instead, focus your memory budget on:
- Your role and organizational context. Title, company, team size, industry, primary stakeholders.
- Active projects. Current top 2-3 priorities with one-sentence descriptions.
- Recurring instructions. Format preferences, terminology standards, communication style.
- Key constraints. Budget limits, compliance requirements, technology stack choices.
Review and prune regularly: go to Settings, then Personalization, then Manage Memories. Delete anything outdated or low-value. Think of it like managing a 1,750-word executive summary of who you are and what you're working on.
Also note the "Reference Chat History" feature launched in April 2025, which lets ChatGPT pull from your past conversations to inform current ones. This helps, but it's still constrained by the same context window limits. The AI can reference past chats, but it's selecting and summarizing, not replaying them in full.
Strategies 1 through 4 are progressively more effective, but they all share a fundamental limitation: they work within the AI platform's constraints. Memory caps, context window resets, and summarization losses are inherent to the architecture. Strategy 5 addresses the architecture itself.
Strategy 5: Build an external knowledge base with MCP
This is the structural fix. Instead of fighting the AI's memory limitations, you bypass them entirely.
The concept: maintain your important context in documents that live outside of ChatGPT. Your company overview, product roadmap, meeting notes, style guide, and project status all exist as documents that any AI conversation can read on demand. The documents don't disappear between sessions. They don't get summarized or compressed. They don't get wiped by platform bugs. They're your documents, and the AI accesses them when it needs them.
The technology that makes this practical is MCP (Model Context Protocol), an open standard for connecting AI tools to external services. With MCP, an AI assistant can read your documents, search through them, update them, and even create new ones based on your instructions.
Here's what this looks like in practice:
You start a new ChatGPT conversation. Instead of re-explaining your company for the hundredth time, you say: "Read my company overview and product roadmap, then suggest Q2 priorities based on what shipped in Q1." The AI reads your actual documents, in full, with no summarization, and gives you advice grounded in your real context.
After a meeting, you say: "Update the weekly meeting notes with today's decisions: we're pushing the HIPAA deadline to March 30, and Sarah is taking over the API docs." The AI updates the document, and tomorrow's conversation will see the new information automatically.
Unmarkdown™ provides this capability with a built-in MCP server. Your documents are written in markdown, the same format AI tools natively understand, so there's no translation layer or information loss. The MCP integration gives the AI seven tools for working with your documents: listing, reading, creating, updating, publishing, converting (for Google Docs, Word, Slack, and more), and usage tracking.
Setup takes about two minutes. The integration guide covers the configuration for every major AI client.
The hierarchy
Use all five strategies together. They're not mutually exclusive.
Strategies 1 through 4 are band-aids of increasing effectiveness. Custom instructions give you a persistent baseline. Projects give you file-level context. Structured conversations optimize what the AI retains within a session. Saved Memories preserve a small amount of cross-session context.
Strategy 5 is the structural fix. It doesn't replace the others. It complements them. Your custom instructions tell the AI who you are. Your Projects organize your workflow. Your conversation structure optimizes the session. Your Saved Memories carry forward the essentials. And your external knowledge base provides the full, uncompressed, always-available context that everything else is too limited to hold.
The people who figure this out early, who build a persistent knowledge base and connect it to their AI tools, end up in a different category of productivity. Every document they add makes every future conversation better. Every update they make is immediately available everywhere. The knowledge compounds instead of evaporating.
Your AI doesn't need a bigger context window. It needs access to your actual documents.
