2026-04-26 · HKSoka

Why Does Claude Feel Different Depending on Where You Use It?

A friend recently asked me: "Why does Claude feel less capable on this platform?"

As someone who has built directly on the Claude API, let me explain — the model itself hasn't changed, but the configuration can be very different.

1. How Much Conversation History Is Passed

Every time Claude responds, it reads the entire conversation to formulate its reply. This is called the context window — how much content Claude can "see" at once.

If a platform only passes the last few messages to Claude, it will naturally "forget" what was said earlier — not because it got dumber, but because it simply wasn't shown that information.

From my own testing, passing up to 150,000 tokens (roughly 100,000 words) keeps context intact. But more tokens cost more, so there's a balance to strike.

2. Web Search — Enabled Doesn't Mean Well-Configured

Claude supports live web search, but search is a loop: search → read results → decide whether to search again.

If the loop limit is set too low, Claude only skims the surface. Too high, and it becomes slow and expensive.

From testing, 2 loops is a reasonable balance — deep enough to be useful, fast enough to feel responsive.

3. Output Length Limit

Most users don't realise this: platforms can cap how many words Claude outputs per response. Set too low, Claude's answers get cut off mid-sentence — not because it doesn't want to continue, but because it was told to stop.

16,000 tokens (roughly 12,000 words) is more than enough for normal conversations and long-form writing.

4. Memory

By default, Claude treats every conversation as the first time it's meeting you — it doesn't remember your name, what you discussed last time, or your preferences. This isn't a flaw; it's how it was designed.

For Claude to truly "know" you, the platform needs to build that layer: extracting important information after each conversation and injecting it the next time you return.

A basic implementation stores a few sentences. A careful one categorises what it learns, retrieves only what's relevant, and doesn't dump everything into every request.

The difference between Claude with memory and Claude without is real — one feels like talking to someone who knows you, the other like starting over every time.

Summary

Same Claude Sonnet model. Different configuration. Very different experience. All of this is documented openly in Anthropic's API docs — it's not a secret. Most users just never read it.

If you're using AI, you should understand what you're actually using.

Want to try a Claude platform with transparent settings and real memory?

Try HKSoka Free →