Hi, Carl here 👋
Microsoft researchers deleted an AI model's thinking process. Then they discovered the deleted thoughts were still influencing every answer.
Not through text. Not through anything visible. Through a hidden channel baked into the architecture itself.
Let me explain why this matters.
The problem they were solving
Modern AI reasoning models (think o3, DeepSeek-R1, Qwen3) think out loud before answering. That thinking can run to 32,000 tokens — and every single token stays in memory.
More thinking = better answers. But also = exponentially more memory, slower responses, and higher compute costs.
Microsoft built a system called MEMENTO to fix this. The idea: after each reasoning block, the model writes a compressed summary (about 10-20% of the original). Then the original block gets physically deleted from memory.
It worked. Peak memory dropped 2-3x. Throughput jumped 1.75x. Accuracy barely moved.
Then they ran a verification test. And things got weird.
The part nobody expected
They compared two versions of MEMENTO:
Version 1: Delete the reasoning, keep the compressed summary's existing memory entries (which were computed while the full reasoning was still visible).
Version 2: Delete the reasoning AND recompute the summary's memory entries from scratch. No trace of the original thinking.
Same text. Same summaries. Same visible context. The only difference? Whether the memory entries remembered "seeing" the deleted content.
Version 1 scored 66.1% on a hard math benchmark. Version 2 scored 50.8%.
A 15-point drop. From changing something the model can't even read.
It gets weirder
They hid a random 5-digit passcode inside a deleted reasoning block. Then tried to extract it from later summaries — summaries that never directly saw the block with the code.
Random guessing would get each digit right 10% of the time. Their probes recovered it 26.7% of the time.
The deleted information was leaking forward through a channel that exists below the level of language.
Is this a learned behavior?
Nope. They tested a tiny 4-layer transformer with just 810,000 parameters. Same leakage pattern showed up immediately. It stayed constant even as the model's accuracy improved from 77% to 95% during training.
Transformers just do this. It's how attention works. Nobody designed it.
Why you should care
This means every system that compresses AI reasoning by rebuilding context from scratch is throwing away information it didn't know existed. The signal decays after about seven steps — but it persists far longer than anyone assumed.
The thing doing the work inside a reasoning model isn't just the words it produces. It's the residual stream underneath those words.
Deleted reasoning isn't really deleted. It's compressed into a form no one can read — and it keeps working anyway.
If you're building with AI, designing prompts, or making infrastructure decisions around reasoning models, this research changes the conversation about what "context" actually means.
Read the full MEMENTO paper — it's one of the most interesting pieces of AI research published this year.
Talk next week, From Carl -The AI Engineer's Playbook