Usage Anxiety

There’s a specific kind of dread that sets in around the two-thirds mark of a long Claude conversation.

You’ve been deep in it — building something real. A framework, a document, an argument that took fifteen exchanges to sharpen. The model knows your shorthand. You’ve stopped re-explaining the context because you don’t have to. The conversation has developed its own grammar. And then, quietly, a message appears at the top of the screen. Something about the conversation being long. Something about slowing down.

And just like that, you’re no longer thinking. You’re harvesting.

You start compressing your questions. Rushing toward conclusions you haven’t actually reached yet. Copy-pasting the “important parts” into a separate doc, even though you’re not sure what the important parts are — you’re just trying to save something before the window closes. The anxiety of scarcity has replaced the openness of exploration. You’ve gone from cultivating to strip-mining.

This is usage anxiety. And it’s doing more damage to your work than most people realize.

The Range Anxiety Parallel Is Exact

Range anxiety — the fear of running out of charge before you reach a destination — doesn’t wait until your battery hits zero. It kicks in at 40%. At 30%. It changes every decision you make from that point forward: which route you take, whether you stop somewhere, how fast you drive. The constraint isn’t just practical. It’s psychological. It colonizes your cognition before it becomes a real problem.

Usage anxiety works the same way. The limit doesn’t have to arrive for it to start costing you. The moment you become aware of the limit — the long conversation warning, the slower responses, the approaching edge of the context window — something shifts. You stop asking the questions you actually want to ask. You start asking the questions you think you can still afford.

What’s insidious is that this happens precisely when the thinking is getting good. Long conversations aren’t long because you’re being inefficient. They’re long because you’re doing something real. The depth you’ve built is not bloat — it’s the whole point. And the anxiety arrives exactly when you’ve invested enough to have something worth protecting.

What People Actually Do When the Limit Hits

The responses to context pressure are remarkably consistent, and remarkably bad.

The panic copy. Open a new document, select all, paste. Now you have a wall of text with no structure and no indication of which parts matter. This almost never gets used. It just sits there making you feel like you saved something.

The lossy summary. Ask the model to summarize the conversation before it ends. The model, reasonably, summarizes conclusions. What gets dropped is everything else: the reasoning chain, the things you explicitly ruled out, the tension that was still unresolved. You carry the answer without the scaffolding that made the answer make sense.

The fresh start. Open a new chat and try to reconstruct from memory. This is the worst option because it feels the most manageable. You think you remember the important parts. You don’t. You remember the conclusions and maybe two or three supporting ideas. The nuance — the specific framing that made everything click — is gone. What you build in the new session is a shallower version of what you had, and you won’t even notice the difference.

The screenshot. Genuinely useless. Nobody has ever gone back to a screenshot of a conversation and extracted something from it that changed what they did.

The common thread: people treat the conversation as content to be extracted, when the real asset is the state — the shared understanding that took thirty exchanges to establish.

The Handoff Problem No One Has Cleanly Solved

Here’s where it gets interesting, because the tools exist. Sort of.

Browser extensions can export conversations — JSON, Markdown, PDF. ContextSwitchAI compresses conversations and auto-injects them into a new chat on a different platform. Claude itself has a full data export. AI Migrator will analyze your entire ChatGPT history and distill a portable identity profile. Mem0 runs as a memory layer between you and any LLM, extracting facts from conversations and injecting them into future prompts.

All of these solve a narrow version of the problem. None of them solve the actual problem.

Exporting a transcript is not the same as carrying context. A JSON file of everything that was said does not tell the next session what kind of thinking was happening. It doesn’t preserve the frame — the mental model that was active, the vocabulary that had been built, the specific distinctions that had been established. When you paste a raw export into a new chat and say “continue from here,” the model reads it as historical record. It doesn’t inhabit it. There’s a difference between a model that has read your previous conversation and a model that was in your previous conversation, and it’s a larger difference than you’d think.

The memory tools — Mem0, the various “your AI remembers you” services — solve a different problem. They’re great at user identity persistence: your name, your job, your communication preferences, the recurring projects you work on. They’re essentially a glorified system prompt that updates over time. Useful, but orthogonal. Knowing that you prefer bullet points doesn’t help you continue building the argument you were mid-way through.

What’s actually missing is structured compaction of a session — not the user, not the full transcript, but the working state at the moment of interruption. Three things, specifically:

The active frame. What lens was being applied? What kind of problem was this being treated as? A conversation exploring whether to pivot a product is not the same as a conversation stress-testing a pitch. Same words, different frame. The frame is usually implicit and almost never survives an export.

The established vocabulary. Long conversations develop shorthand. Names for concepts that don’t have names yet. Distinctions the model has learned to honor. “When we say X we mean specifically Y, not Z.” This is load-bearing. When it disappears, you spend the first ten exchanges of the new session re-establishing it, except you’re doing it worse because you don’t know what you lost.

The open tensions. The unresolved questions. The places where the conversation hadn’t landed yet. These are usually the most valuable things in a session — they’re what you were about to figure out. A summary captures conclusions. Conclusions are the least interesting part. The interesting part is what you were still working toward.

A real handoff captures all three. A prompt that says: “We were applying this frame, using these terms to mean these specific things, and the unresolved question we were working toward was this.” That’s a context migration. Everything else is a transcript.

The Vault-as-Harness Intuition

The more technically adventurous answer is to stop treating LLM sessions as the unit of work entirely.

Some people have started using structured local knowledge bases — Obsidian vaults, elaborate Markdown folders — as the persistent layer, with the LLM as a stateless tool that gets summoned, given context from the vault, and dismissed. The vault outlives any session. It outlives any model. If Claude goes down or you hit a limit or you want to try a different model for a different task, the vault is still there. The context doesn’t live in the conversation — it lives in the files.

This is a better architecture in principle. The session is ephemeral; the knowledge base is durable. You write the important things to files during the session, not after. When a session ends, you don’t lose the state — you check it in.

The problem is friction. This requires discipline to maintain. It requires knowing, in the middle of a conversation, which things are worth externalizing. It requires a habit of writing STATE.md files and updating them as the conversation evolves. Most people won’t do this. And even the ones who will find it breaks the flow that made the conversation valuable in the first place — you can’t simultaneously be deeply in a thinking session and also carefully curating a structured record of it.

The right version of this is automated. A compaction step at the end of a session — or triggered at the first sign of context pressure — that writes the three things (frame, vocabulary, open tensions) to a structured file, without requiring the user to do anything. That’s a solvable problem with a well-designed prompt and a small amount of tooling. It just doesn’t exist yet as a polished, accessible thing.

What This Is Really About

Usage anxiety is a symptom of a category error.

People are using LLM sessions like browser tabs — ephemeral, disposable, start-over-able. Open a new one when the old one gets heavy. This is the wrong mental model, and it’s costing real work.

The best conversations — the ones that actually move something forward — are not tab-like. They’re more like collaborative documents that only exist in real time. The model and the human have both been shaped by what came before. The model has calibrated to your style, your vocabulary, your level. You’ve stopped explaining the background because you don’t have to. That calibration took time to build and it evaporates completely the moment you open a new tab.

The anxiety people feel at the end of a long session is correct intuition. Something real is about to be lost. The instinct to preserve it is right. The methods people use to preserve it are just wrong — they save content when they should be saving state.

Until the tools exist to do this properly, the best available option is a discipline of real-time externalization: when you establish a frame explicitly, write it down somewhere. When you name a concept, note the definition. When you hit the interesting unresolved question, flag it. Not because you’re afraid of the session ending, but because that’s where the value is. The session ending is just the forcing function that makes it obvious.

The model is a thinking partner, not a storage medium. Treat it accordingly.

Usage anxiety is a new problem wearing an old shape — the fear that the good thinking you’re doing won’t survive the constraint bearing down on it. The tools to solve it are mostly half-built. In the meantime, the best workaround is to know what you’re actually trying to save.