Two numbers travel with every write-up of Anthropic's context-management features, including ours. Context editing cut token consumption by 84% on a 100-turn web-search evaluation. Pairing the memory tool with context editing lifted task success by 39% on complex multi-step benchmarks. They are real, they come from the vendor's own announcement, and they are quoted as if they mean the model got better.

They don't mean that. Both numbers measure the same thing from two angles: what it's worth to stop hitting the context wall. Whether either one applies to your agent is a question about your workload, not about Claude — and the answer is often "less than you'd hope," sometimes "zero," and occasionally "more than 84%."

What the 84% is a ratio against#

The load-bearing detail is buried in the same sentence as the number: context editing let agents complete workflows that would otherwise fail due to context exhaustion. The baseline wasn't a healthy run using more tokens. It was a run drowning — filling its window with stale tool output until the signal was buried and the task collapsed. The 84% is how much of that waste the clearing removed.

So the number is a property of a specific failure mode. Context editing waits until your input crosses a trigger (default around 100,000 tokens), then clears the oldest tool results while keeping the most recent few, replacing each with a placeholder the model can re-fetch. On a 100-turn agent that re-reads files and re-runs searches, the window fills with re-fetchable junk, and clearing it is nearly free upside — hence a very large ratio.

Now run a single-shot summarizer, or a classifier, or any agent whose traffic never approaches 100,000 tokens. Nothing crosses the trigger. Nothing clears. There is no 84% to capture, because there was no wall to escape. The number didn't shrink for your workload — it was never there.

The 84% doesn't describe a smarter model. It describes how much stale tool output a long agent was carrying before something threw it away.

What the 39% actually unlocks#

The 39% is the same story in success-rate units. Add the memory tool — a scratch filesystem Claude reads and writes so facts survive a context reset — on top of context editing, and complex multi-step tasks complete 39% more often than with neither.

Again: complex multi-step. The gain lives on tasks long enough that state falls out of the window and has to be reloaded. The memory tool's whole value is persistence across a reset; a task that never resets has nothing for it to persist. On short work the 39% doesn't shrink gracefully to 30% or 20% — it collapses toward zero, because you've paid for a mechanism the task never exercises.

This is the trap in both figures. They're not dials that scale smoothly with effort. They're step functions keyed to whether your agent hits the context wall. Above the wall, large. Below it, absent.

The cost the numbers don't print#

There's a line item neither headline includes: your prompt cache. Every clearing event rewrites the cached prefix of your prompt, which invalidates previously cached tokens and forces a re-write on the next call. A clear that frees a few thousand tokens can cost more in cache misses than it saves — which is exactly why Anthropic ships a clear_at_least knob that stops an edit from firing unless it frees enough to pay for itself.

And below the wall, editing can go actively negative. Clear a tool result the model still needed, on a task where re-fetching it is expensive or impossible, and you've traded a token saving for a wrong answer. The feature assumes tool output is cheap to re-derive. When that assumption breaks, so does the free lunch.

The number worth quoting is yours#

None of this makes the features weak. Context editing and the memory tool are the right tools for long-horizon agents — the ones that genuinely exhaust their windows, where 84% and 39% are not marketing but survival. The point is narrower and it's the same one that governs every agent-memory benchmark: a headline number is uninterpretable without the baseline that produced it, and here the baseline is failure.

So don't budget on the announcement. Run your agent with and without editing on a representative long task, and read context_management.applied_edits on each response to see what actually cleared. If your runs rarely cross the trigger, you'll see a small number and should size your expectations to it. If they routinely fill the window and your tool output is re-fetchable, you may beat 84%. Either way the figure you can act on is the one you measured — not the one you inherited.