Documentation
Context Fidelity™
Our benchmark for how much meaning survives a model handoff. It measures what's preserved when context moves Claude → GPT → Gemini, so nothing important is lost in translation.
The problem it measures
Every time context crosses from one model to another, meaning can leak — names get dropped, intent flattens, the reasoning behind a decision disappears. Context Fidelity™ puts a number on that loss so it can be tracked and improved rather than ignored.
What we measure
Rather than counting raw tokens preserved, Context Fidelity focuses on preserved meaning:
- Entities — the people, projects and things a memory refers to.
- Decisions — what was concluded, and why.
- Relationships — how memories connect to one another.
- Intent — what the user was actually trying to do.
The .aal format is what makes a high score achievable: it carries this structure with the memory, so the receiving model inherits context instead of guessing at it.
Reading the score
| Metric | Meaning |
|---|---|
| retained % | Share of meaning preserved after a handoff (target: 98%). |
| chain depth | How many models context passes through intact (3+). |
| shared brain | One memory layer read and written by every model. |
The figures shown across the site are illustrative targets while the benchmark matures. Methodology and live results will be published here.