For years, the common approach to AI tools was the same as meeting a stranger: introduce yourself, explain the context, do the work. The next day — same stranger, same introduction. This seemed normal. It isn’t. It’s a fixable infrastructure problem, and the last few months have gone into fixing it.
The re-explanation tax
Every new chat window carries a hidden cost. Four minutes here, six minutes there, rebuilding the context the AI needs to actually help: what’s in progress, who the stakeholders are, what’s already been tried. Then the session closes, and the cost repeats tomorrow.
This isn’t a prompting problem. The best prompt in the world can’t compensate for an AI that doesn’t know what you’ve been working on.
It’s also not a single-tool problem. Claude’s memory is separate from ChatGPT’s memory, which is separate from Cursor’s context. Five sticky notes on five separate desks. Switch tools — and you will switch, because different models genuinely are better at different tasks — and you start over.
Why the obvious solutions don’t quite work
The obvious approaches are ones most people try. One approach: paste a big context block into the system prompt. Another: keep a Notion page to copy from. Another: rely on built-in memory features. Each approach held for a while, then broke: the system prompt went stale, the Notion page fell behind, the built-in memory didn’t carry across tools.
The structural problem is that tools like Notion and Obsidian were designed for human eyes — browseable, beautifully laid out, searchable by keyword. They’re excellent for what they do. But an AI agent can’t browse. It needs a protocol, a structured interface it can call programmatically. Bolting a chatbot onto a human-interface app just trades one silo for another.
What actually changed things
The approach that stuck was treating context as infrastructure rather than as a per-session chore.
The core idea is what might be called an “Open Brain”: a database-backed memory store that any AI tool can query. The implementation uses Postgres with pgvector — that’s a standard database with an extension for vector search, which means it can find notes by meaning rather than just by exact keyword match — connected via MCP, an open protocol that Claude, ChatGPT, Cursor, and a growing list of other tools all support.
When something gets captured — a decision, a pattern, a lesson from a difficult meeting — it goes into that store. Retrieval is semantic: a query like “what was the thinking on the product roadmap last month?” surfaces the relevant note even if those exact words never appeared. The system makes a best guess at metadata classification, and sometimes it’s off, but the meaning-based search compensates. In practice, it finds what’s needed.
The practical cost is around ten cents a month. The setup, following a step-by-step guide, takes about 45 minutes. No code required beyond following instructions.
The compounding effect
What makes this worth writing about isn’t the setup — it’s what accumulates afterward.
Context infrastructure compounds in a way that better prompts don’t. Every captured thought makes the next interaction more precise. Every week that passes, the difference between working with this system and working without it widens. It’s not additive. It’s cumulative.
This is mostly a question of switching cost. A colleague who routes different tasks to different models — one for reasoning, another for writing, another for code — can only do that fluidly if context moves with them. If every tool switch means rebuilding context, the cost is prohibitive and people default to one tool out of friction, not preference. If context is centralized and any tool can query it, the switching cost drops to nearly zero. That’s when you can genuinely use the best tool for each job.
The people who are most effective with AI right now aren’t necessarily writing better prompts. They’ve built better context infrastructure.
Where this is going
The infrastructure layer for AI work is still being assembled. MCP, the protocol that makes tools like the Open Brain possible, started as an Anthropic experiment and is now backed by OpenAI, Google, Microsoft, and the Linux Foundation. That trajectory suggests it’s becoming actual infrastructure — the way HTTP is infrastructure — rather than any single company’s proprietary bet.
That matters for how we build. A system on open primitives — a real database, an open protocol — is one we control. We’re not locked into one platform’s memory, not hostage to a SaaS company’s next pricing decision.
The gap between occasional AI use and AI embedded in how one thinks and works turns out to be mostly an infrastructure question. The good news: infrastructure is buildable, even if you’re not an engineer. Forty-five minutes and a willingness to capture thoughts consistently is the real investment. The compounding starts immediately.
Text summarized and optimized using Anthropic’s models and reviewed by a human.