Model upgrades are routine, yet memory migrations receive little attention. An agent may keep the same memory store but a new model can reinterpret old notes, mixed embedding versions can break retrieval, and repair may fail without the original evidence.
This study compares four memory representations: long‑context raw reading (LC-RAW), chunked retrieval‑augmented generation (RAG), model‑compressed natural‑language notes (NOTES), and a fixed‑schema knowledge graph (KG-fixed).
We evaluate 48 synthetic histories with randomized answer codes, exact scoring, and two open‑weight models under 10 billion parameters.
Results show that fixed‑schema structures transfer reliably, with KG-fixed accuracy changing by only $+0.0004 \pm 0.0020$. In contrast, compressed NOTES exhibit strong model coupling, with accuracy shifting asymmetrically by $+9.91$ or $-13.28$ percentage points depending on migration direction.
In RAG systems, a 50/50 mixed embedding index yields only a 4.96‑point accuracy gain, far short of the 11.90‑point improvement achieved through full re‑embedding.
Diagnostic decomposition attributes 80% ($0.467 \pm 0.014$) of the NOTES accuracy deficit to information lost during initial construction, while retrieval failures account for 81% ($0.364 \pm 0.012$) of the RAG deficit.
Store‑only repair of NOTES fails to reach a 90% performance recovery target across all 48 cases; retaining the raw source history enables successful recovery in 34 of 48 cases for one migration direction.
These findings highlight the need for direction‑specific migration testing, strict embedding space isolation, and retention of source histories to support memory repair.
Review