T4. Cross-model handoff: the shared brain
Every other benchmark hand-seeds the memory pool. T4 exercises the harder half of the loop, capture, end to end, and across a model boundary.
- Model A (opus) invents five arbitrary, non-derivable project decisions in a real session (a service name, port, retry limit, cache TTL, auth token prefix, values a second model cannot guess).
- Those decisions are captured through the real pipeline into an isolated pool.
- Model B (sonnet), a different model in a fresh session with no shared history, is asked to write code that requires those decisions, with the retrieved memory injected. A blinded B (no memory) is the control.
Grading is objective: how many of A's exact values appear in B's output.
| Arm | A's decisions honored |
|---|---|
| Memory-B (recall injected) | 1.00 |
| Blinded-B (no memory) | 0.00 |
Memory carried A's invented values across both a model and a session boundary; the blinded model correctly reported it had no record of them. The captured decision was confirmed retrievable, which also validates the capture quality gate: it keeps genuine decisions and only drops non-memories.
Honest note. An early run showed a "blinded" model scoring perfectly, which is impossible by guessing. The cause was a harness bug: the model was running where it could read the test's own source files. It was fixed by running each session in an empty working directory with tools disabled, so blinded is truly blind, which also matches reality: a downstream model sees only what memory carries, not the upstream session's files. The result above is post-fix.