The Context Crisis: Solving the LLM Memory Problem in High-Scale GTM Strategy
TL;DR: Every senior PMM who has used an LLM for GTM work has hit the same wall: the output is technically competent but strategically shallow. The model doesn’t know your ICP deeply enough. It doesn’t know how your positioning has evolved. It doesn’t remember the decision you made last quarter. This is the Context Crisis, and it’s entirely solvable — if you’re willing to architect for it instead of prompting around it.
The Real Bottleneck Is Not the Model
The AI marketing conversation is dominated by model comparisons. Which LLM writes the best copy? Which one handles long-form better? Which one understands B2B positioning?
These are the wrong questions. The frontier models — Claude, GPT-4, Gemini — are all genuinely capable of producing sophisticated GTM content. The problem is not capability. The problem is that capability without context produces work that is technically correct and strategically empty.
If you ask any frontier model to write a positioning statement for your product, you will get a positioning statement. It will follow the correct structure. It will use appropriate B2B language. It will be utterly, catastrophically generic. Because the model does not know your customer the way you know your customer. It does not know the specific anxieties of your ICP. It does not know the competitive narrative shift you made eighteen months ago and why. It does not know that the phrase “enterprise-grade” is specifically off-limits because your last VP of Marketing tested it and it tanked conversion.
All of that knowledge lives in your head, in scattered documents, and in the institutional memory of a team that has partially turned over. The Context Crisis is what happens when you try to apply AI capability to a problem without solving the context problem first.
What Persistent Context Architecture Actually Means
The solution is not better prompting. Better prompting is a workaround for missing context. It treats every interaction as a fresh start and puts the burden on you to manually reconstruct the relevant background every single time.
Persistent context architecture means building a structured, maintained set of documents that function as the permanent operating context for every AI-assisted GTM task. Think of it as the difference between hiring a consultant who reads a one-page brief before every meeting and hiring someone who has been embedded with your team for a year. The output quality difference is not marginal. It is categorical.
For a PMM Lead, the core context corpus typically has three layers.
Brand Context covers voice, tone, stylistic rules, terminology decisions, and the explicit list of things you never say and why. This is not a generic brand guide. It is a living document that captures the reasoning behind every significant communication decision.
ICP Context goes beyond the standard demographic and firmographic profile. It captures the specific anxieties that drive purchase decisions in your category, the language your buyers use that is distinct from the language your company uses, the objections that keep appearing in sales calls, and the moments of genuine competitive differentiation that resonate versus the ones that land flat.
Positioning Context documents your current strategic narrative, how it has evolved, what was tried and discarded, and the specific competitive claims you are making and the evidence behind them. This layer is the one most companies have never actually written down coherently.
Claude Code’s Project Memory as a Model for GTM Context
Claude Code’s project memory feature is one of the clearest practical demonstrations of what persistent context architecture looks like in a professional workflow. When you configure Claude Code with a CLAUDE.md file in your project root, every subsequent interaction in that project benefits from that accumulated context. The model knows the project conventions, the technical constraints, the decisions that have already been made.
The same principle applies directly to GTM work. A CONTEXT.md file — or a structured folder of context documents — that sits at the root of your GTM repository becomes the living memory of your marketing strategy. Every AI-assisted task runs against that context. The model drafting your launch email knows your ICP’s primary anxiety. The model writing your competitive battlecard knows the claims you have evidence for and the ones you don’t.
This is not a hypothetical future state. It requires disciplined document maintenance and a commit to treating your context corpus as a first-class asset — not a one-time setup that you forget about.
Structuring Context for Agentic Workflows at Scale
When you move beyond individual AI interactions into full Agentic Workflows — multi-step pipelines where agents are completing tasks autonomously and passing outputs to subsequent steps — context architecture becomes even more critical.
An agentic campaign pipeline might involve an agent that monitors competitive activity, another that synthesizes that intelligence into a positioning brief, another that drafts content based on that brief, and a final review step before human approval. For this pipeline to produce strategically coherent output, every agent in the chain needs access to the same core context. The competitive monitoring agent needs to know what signals actually matter for your positioning. The content drafting agent needs to know your voice, your ICP, and your current narrative emphasis.
This is solved by treating your context corpus as a shared system prompt layer that every agent in the pipeline inherits. In practice, this means your context documents are stored in your Git repository alongside your content, versioned with the same discipline as your GTM assets, and injected into every agentic task that requires strategic judgment.
The result is a pipeline that produces work that is not just technically competent but genuinely aligned with your strategy. Campaigns that accumulate rather than reset. An AI workflow that gets more strategically coherent over time rather than starting fresh with every prompt.
The Maintenance Discipline
Persistent context architecture fails without one thing: the discipline to maintain the documents. Most teams build the initial context corpus with good intentions and then let it drift out of date. Six months later, the ICP document reflects a buyer profile the team has since refined significantly. The positioning context still references a competitive narrative that has been superseded. The brand context doesn’t include the terminology decisions made during the last rebrand.
Outdated context is, in some ways, worse than no context. It confidently produces output that is wrong in ways that are harder to catch than generic mediocrity.
The fix is treating context maintenance as a regular GTM cadence — not a project. Quarterly context reviews that update each layer. A pull request process for significant positioning changes that updates the context corpus as part of the approval workflow. New team members who are onboarded to the context corpus as their first task, which also serves as a forcing function to identify gaps.
Key Takeaway for PMM Leads
The frontier models are not your bottleneck. Your context architecture is. Building a structured, versioned context corpus — covering brand voice, ICP depth, and positioning evolution — is the highest-leverage investment you can make in your AI-assisted GTM capability. It is the difference between a tool that produces generic competence and a system that compounds your strategic thinking. Start with ICP context. Get specific. Maintain it with the same rigor you bring to your positioning itself.
FAQ
Q: How long does it take to build an initial context corpus? A meaningful first version — one that genuinely improves AI output quality — can be built in a concentrated two-day session. The ICP context layer typically takes longest because most teams discover they have less documented specificity than they thought. The brand context layer is usually faster. Budget a week for a complete first version with proper review.
Q: Should context documents be public or private? The context corpus contains sensitive competitive positioning and ICP intelligence. It should live in a private repository. If you are using a personal site built on Astro 6.0 with a public GitHub repository, keep your GTM context corpus in a separate private repo and inject it via environment variables or a protected API layer.
Q: How do I handle context for multiple products or market segments? Build a hierarchical structure: a core company-level context layer that all products inherit, with product-specific layers that override or extend the core where needed. The same pattern engineers use for environment configuration translates cleanly to GTM context management.
Q: What if my context documents become too long for a single system prompt? This is a real constraint and requires a retrieval architecture — essentially a semantic search layer that pulls the most relevant context sections for each specific task rather than injecting the entire corpus. For most PMM Leads, this is not necessary until the corpus exceeds roughly 50,000 words. Start with the full injection approach and add retrieval when you hit the limit.