The memory layer is the moat

I use Claude Code, Claude Desktop, Claude Co-Work, Google’s CLI, Gemini, OpenAI’s Codex, and ChatGPT in a typical week. Until recently, none of them knew what the others had been told. Every session started cold, and every tool was an island, which meant I was re-explaining the same context over and over again in every tool, every session, every day.

So I built a persistent memory layer that sits outside all of them. It has its own MCP, storage, and retrieval logic, along with a knowledge graph that any tool can query. It took a few weeks to get right, and I think it’s the most consequential thing I’ve built this year.

What it actually does

The memory layer stores structured facts, freeform notes, and a knowledge graph with entities, relationships, and semantic search. Every tool I use connects to it on every request, so when I tell one tool something, every other tool knows it. When I make a decision in one context, that decision carries forward everywhere.

What this eliminates in practice is the re-explanation tax that most people pay without noticing. The same preferences, the same project history, the same prior decisions, all re-entered across every session and every tool. Last week, I asked one tool to prepare a project brief, and it pulled my preferences, prior decisions, and relevant notes from across three months of work without me pointing it at anything. That is the difference between a tool that remembers and a tool that doesn’t.

Memory is the real lock-in

I think the conversation about AI vendor lock-in mostly focuses on the wrong things. Model quality, pricing tiers, and feature sets get the attention, but they are not what actually bind you to a provider. The binding mechanism is memory.

OpenAI’s memory, Claude’s project knowledge, and Google’s context caching are all designed to accumulate your information inside one ecosystem. Spend six months building up context inside any of them, and switching providers means abandoning all of it. The models themselves are increasingly interchangeable, but the accumulated context is not, and I think that asymmetry will define the next few years of the market more than any benchmark improvement will.

Building the memory layer externally breaks that dependency entirely. I can swap models without losing context, and I can use the best tool for each task without fragmenting my knowledge across five different systems. The memory belongs to me, not to any provider, and that is worth more than any single model improvement.

What happened when I connected it to Google Drive

The most instructive example of what this enables came when I connected Claude Co-Work to my Google Drive through the memory layer. It scanned my entire Drive, indexed the contents, created a schema and taxonomy for the information it found, and ported the whole structured index back into my persistent memory layer. Not a summary but a working knowledge architecture that I can now query, reference, and build on.

I did not sit there tagging files or building folder hierarchies. The system looked at what I had, figured out how it related, and organised it. That is a task that would have taken me days to do manually, and I know I would never have maintained it. The memory layer made it possible because the system already had enough context about my work and my preferences to make sensible decisions about how to structure what it found.

What changes when context persists

The gap between persistent context and disposable context is larger than most people realise, and it compounds over time. In a disposable context, every interaction is a standalone event in which the tool knows nothing about your work, your preferences, your prior decisions, or the relationships among the things you have asked it to do. With persistent context, every fact stored and every relationship mapped makes every subsequent interaction marginally better.

I think this is where the real leverage in AI productivity lives, and it is mostly absent from the conversation. The discourse focuses on prompt technique and model capability, both of which matter, but neither of which addresses the fundamental constraint that most AI tools forget everything between sessions. Some platforms can now query past chats to pull fragments into the context window, which helps, but retrieving a snippet from a conversation you had three weeks ago is not the same as structured memory that works across tools and compounds over time. A mediocre model with excellent context will frequently outperform an excellent model with no context, because the quality of the output is bounded by the quality of the input, and context is the input that matters most.

The organisations and individuals who figure out how to give their tools a shared, persistent context will have a structural advantage over those who don’t. Not because the technology is hard to build, but because the accumulated context itself becomes irreplaceable. The tools will keep getting better. The memory is what makes them yours.