You upgraded the model.
The notes were written for the old one.

Four ways to keep the same history, and only one of them needs a model

An agent's memory is not a database bolted to the side of it. It is a boundary between past experience and future decisions, and every format loses something crossing it: raw history costs context, retrieval can miss the record, notes can drop evidence while compressing, a fixed schema can only hold the facts it was designed for.

The study builds all four from the identical 48 histories — 160 probes each covering direct facts, changes over time, contradictions, multi-step relations and aliases, with answer codes randomized so no model can recall them from pretraining, and exact-match scoring with no model acting as judge. Step through the formats and watch what each reader scores on the store built for itself.

Two of these four are settled before any upgrade happens. Raw history and retrieved chunks hold the transcript verbatim, so nothing about them changes when the reader changes. The other two are written by a model — and only one of those writes into a structure the next model does not have to interpret.

The same migration, run both ways, has opposite signs

Now change the reader. Each model builds its own notes store and its own schema store, then reads both — its own, and the one the other model wrote. Nothing else moves: same histories, same probes, same scoring.

The filled dot is the reader on its own store; the ring is the same reader on memory inherited from the other model. Step through the four cells.

The asymmetry is not a byte budget. A calibration sweep up to 160 KiB found the stronger writer retaining 85.6% of the required evidence in 143 KiB, while the weaker writer retained 64.5% using 159 KiB — more space, less evidence. What the note-writing model can do is part of the memory design, not a detail of it.

Which is why the paper's own pre-registered summary of the swap looks like a null result. The two planned contrasts average the two directions into one symmetric penalty, and a −13.28 and a +9.91 largely cancel. Both came back below the study's five-point action threshold, while the two tests that did clear it are about index migration and repair.

One format loses the evidence writing it down, the other loses it looking it up

Knowing a store is worse is useless without knowing which stage broke. The study supplies each probe three ways — normal reading, the exact stored items that contain the answer, and the original raw event — and the differences split total error into three additive channels: search inside the store, the writing of the store, and the reader itself.

Notes lose 80% of their deficit at construction — in the worst direction, 88.8%. The content is gone before anyone reads it, which is why rewriting the notes in a friendlier style buys nothing: one model rewriting the other's notes moved accuracy by −0.012, the reverse by −0.042, and the weaker rewriter dropped exact identifiers in 46% of its rewrites. Retrieval takes 81% of the RAG deficit instead: supplying the correct chunks lifts accuracy from 0.53–0.56 to 0.88–0.95, so the chunks were fine and the search picked the wrong ones. Before you blame a new model for forgetting, check that the evidence reached it at all.

Both embedding versions return 1,024 numbers, so nothing errors

The other routine upgrade is the embedding model. Teams adopt a better one and, sensibly enough, re-embed as they go — new writes with the new model, old vectors left alone until there is time. This section is entirely inside the mechanical layer, so both bars are the same colour; the point is that being mechanical does not make a migration safe when it is only half done.

A full re-embed is worth 11.90 points; the half-migrated index captures 4.96 of them and gives back nearly sixty per cent of the upgrade. Old and new vectors are both 1,024-dimensional, so no dimension check fires, no exception is raised, and the system keeps serving — it simply ranks across two incompatible spaces and quietly loses the experiences stored in the older one. Build the new index separately, cut over whole, and keep the two spaces apart.

Price the repair before you need it illustrative

The last question is the one that decides your design: when a migration does go wrong, what does it cost to get the memory back? The study sets a recovery target — 90% of the accuracy a store built fresh for the new model would give — and tries four repairs on every history.

Rewriting notes from the notes reaches the target in none of the 48 cases, in either direction, because a rewrite cannot restore evidence the first compression already dropped. Rebuilding from a retained raw history reaches it in 34 of 48 — but only when the repairing model is strong enough to finish the job; the weaker one ran out of output tokens on every history. Mechanical rebuilds are close to free.

What the paper actually reports