In March 2026, Anthropic removed its long-context surcharge, making Claude's full 1-million-token window available at standard per-token pricing. Google, OpenAI, and several open-weight labs have made comparable moves.
For a lot of people, this landed with a tempting conclusion attached: if the context window is this big and this cheap, do I still need a memory tool?
It's a fair question, and it rests on a mix-up that's easy to make — treating a bigger context window as the same thing as memory. They aren't. If you opened a new Claude or ChatGPT chat this week and had to re-explain your project from scratch, you already know the 1M-token window didn't fix that.
What actually changed in 2026
The numbers are genuinely impressive. The 1M-token tier is now crowded: GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro, Qwen 3.6 Plus, and Llama 4 Maverick all sit at that mark, with Llama 4 Scout pushing to 10 million. Most providers now offer flat-rate pricing across the full window rather than charging a premium past a threshold.
In practice, you can paste in a large document, a long codebase, or hours of transcript, and the model can technically see all of it in one request. That's genuinely useful for one-off, single-session work.
But none of that is memory. It's working space for a single conversation, and it disappears the moment that conversation ends.
So why does Claude still forget you tomorrow?
A context window resets with every new chat. Even a 1M-token one only holds what's inside the current conversation. Close the tab, start a new chat, and it's gone. The model isn't choosing to forget you — there's nothing left to remember from. A 4K window and a 1M window both go to zero the second a new session starts.
Cost still scales with every message. Removing the surcharge doesn't make long conversations free. Every API call reprocesses the entire context you send: system prompt, instructions, full history, documents. A chat that starts at a couple thousand tokens can balloon past 25,000 by message twenty, and you pay for that full amount on every single turn.
"Lost in the middle" is still real. Even inside one large window, models don't treat every token equally. Independent benchmarks — NVIDIA's RULER, Adobe's NoLiMa, Chroma's context-rot research — consistently find that effective context lags advertised context, with recall degrading for information buried mid-prompt. Newer models have improved here (Opus 4.6 scored 76% on MRCR v2 against 18.5% for its predecessor), but degradation still occurs at every length increment. The practical takeaway from that research is the same one memory systems are built on: send relevant content rather than maximizing volume.
One context window, one platform. This is the one a bigger window can never solve. Everything you build up in a Claude conversation stays in Claude. Switch to ChatGPT for a different task and you're starting from zero — not because of window size, but because context windows don't travel between platforms.
Context window vs. memory
Think of a context window as RAM, not storage.
RAM is fast working space for the task in front of you — and it clears when the session ends. Storage is what persists: the facts, preferences, and history that should still be there next week, in a different chat, on a different AI entirely.
A bigger window makes the RAM roomier. It does nothing for storage.
These are two different engineering problems, and the 2026 context-window race has only solved one of them.
What actually solves the memory problem
The fix isn't a bigger window inside one AI. It's a memory layer that sits outside all of them — you store context once, in a place that isn't tied to any single model or vendor.
Four things worth looking for when evaluating one:
**Persistent** — memories survive closed tabs, new sessions, months of time.
**Cross-platform** — works whether you're in Claude, ChatGPT, or Gemini.
**Selective retrieval** — pulls only what's relevant, not your entire history.
**Yours** — encrypted, deletable, not vendor-gatekept.
This is exactly the gap the Model Context Protocol (MCP) was built to close. MCP is now supported across Claude and ChatGPT, which means a single memory server can plug into whichever AI you're using and hand it precisely the relevant context — without you repeating yourself or paying to resend everything.
That's the approach behind Lumi: store wide, retrieve narrow. Everything is kept; only what's relevant to the current prompt surfaces.
The real fix
Context windows will keep growing, and they're worth using well. Use the 1M window for what it's actually good at — large unfamiliar codebases, cross-file reasoning, single-session document analysis.
But if your real problem is repeating yourself every morning, the fix was never a bigger window. It's a memory layer that doesn't reset when the chat does.