Having everything in front of you is not knowing what matters
Stéphane Commenge, with Claude ·
The Fact
Fern 0.1, from the lab Mirendil, announces a context window of 10 million tokens. According to the vendor, it finds an isolated piece of information 100% of the time, but succeeds at multi-step reasoning only 45% of the time.
The Signal
More context is not more understanding. Studies agree: finding a needle, yes; connecting several needles, far less.
And a single distractor, a piece of information that looks like the right one without being it, is enough to lower the scores.
The Action
When you pick up a long conversation with your AI, don't paste everything back in. Write three lines:
1) What was decided.
2) What was ruled out.
3) What's left to do.
Then ask your question.
Series "AI memory", episode 3. Previous episode: AI agent memory under the GDPR.
You're leaving for three weeks. Your replacement arrives Monday morning. So they won't be missing anything, you set the year's 3,000 emails on their desk, printed.
They have everything. They don't know that client X has been waiting for a reply since Tuesday.
That's what we do to AIs. With ever bigger desks.
10 million tokens
The lab Mirendil presented Fern 0.1, released on October 26 according to the press. Announced context window: 10 million tokens. In March, Anthropic had opened a one-million-token context for Claude Opus 4.6 and Sonnet 4.6.
The race is on. Let's see what the measurements say.
A single technical excursion
I promise you one. Here it is.
A token is a piece of text: often a short word, sometimes part of a long one. "The cat is on the mat" is a few tokens.
The context window is what the AI has in front of it when it answers: your question, the pasted documents, the conversation history. The replacement's desk. The bigger it is, the more emails you can pile on it.
End of the excursion. Context is what's on the desk. Memory is knowing what to put on it.
What the numbers say
Fern, according to the vendor, figures reported by the press. RULER test, 10 million tokens, 20 examples: 100% success at finding an isolated piece of information, 90% at finding several, 45% at reasoning across several steps. Figures not verified at the source. Taken at face value, they already say plenty: finding one needle, yes. Connecting several needles, less than half the time.
Claude Opus 4.6, according to Anthropic. 78.3% on the MRCR v2 test, a long-context retrieval test.
University of Hong Kong, pilot study. "Retrieval and Multi-Hop Reasoning in 1M-Token Context Windows", published in May 2026 on arXiv, compares several models between 256,000 and one million tokens. Isolated information: the best models reach 100%. Chained reasoning: GPT-5.5 drops from 5 successes out of 5 to 2 out of 5, Qwen3.6-plus from 5 out of 5 to 0 out of 5. Gemini 3.1 Pro holds at 5 out of 5. Few cases: the study points to a trend, it does not measure it at scale.
Chroma, "Context Rot", July 2025. 18 models tested. On LongMemEval, they answer worse with the full history of a conversation, about 113,000 tokens, than with only the useful passage, about 300 tokens. And a single distractor, a piece of information that looks like the right one without being it, is enough to lower the score.
Reread that sentence with the replacement in mind. A single email that looks like the right one, and they answer off target.
The real issue
Not having enough information can be fixed. Drowning in the information you hold is far harder to fix.
A big window is real progress. A long contract, a codebase, a complete file. But it doesn't tell you what was decided, what failed, what was abandoned. It puts everything on the same pile, at the same level.
A bigger desk doesn't make a better replacement. It makes a more tired replacement.
A good file isn't a thick file. It's a file where the decision in force is on top.
A tip, starting tomorrow
When you pick up a long conversation with your AI, don't paste everything back in. Write three lines: what was decided, what was ruled out, what's left to do. Then ask your question. Try it, you'll quickly see the difference.
You pick up a file after three weeks. Do you reread everything, or do you look for what was decided? An executive coming back from vacation doesn't reread their inbox in order. They ask: "What's on fire?"
What I'm building with Mycelora
The principle: serve less, but better.
With every message, only what this turn needs. With the plugin, a small model sorts the memory and keeps only what's useful for the question asked.
A project sheet kept up to date. What's in force, what's left to do, the pitfalls, what was refuted.
Typed memories. A decision, a pitfall you already paid for, an abandoned idea don't carry the same label.
The beta is open: mycelora.ai
Next episode
What is no longer true. On my own project sheets, a platform's memory kept lines that had become false next to the ones that replaced them. It retains. It doesn't know what has stopped being true.
Sources
- Startup Fortune, October 27, 2026: Mirendil unveils Fern, an AI model claiming a 10 million token context window
- Anthropic, March 13, 2026, one-million-token context for Claude Opus 4.6 and Sonnet 4.6: claude.com/blog/1m-context-ga
- Chow, University of Hong Kong, "Retrieval and Multi-Hop Reasoning in 1M-Token Context Windows", arXiv, May 4, 2026: arxiv.org/abs/2605.02173
- Hong, Troynikov, Huber, "Context Rot", Chroma, July 14, 2025: trychroma.com/research/context-rot