# Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory Portability

Source: [arXiv](https://arxiv.org/abs/2609.05339v1)  
Feed7 permalink: https://feed7.dev/p/2609-05339v1-06dws1c  
Published: 2026-09-04T16:44:17.000Z  
Trust: Needs Review (needs_review)

## Why Included

Agent memory can break across model upgrades even when the store is unchanged. Test migrations by direction, rebuild embeddings completely, and retain raw histories for repair.

## Source Summary

Across **48 synthetic histories**, fixed-schema knowledge graphs barely changed after a writer-model swap, while compressed notes moved by **+9.91 or −13.28 percentage points** depending on migration direction. A half-migrated RAG index gained 4.96 points versus **11.90 points with full re-embedding**.

## Practical Implication

Treat model upgrades as memory migrations: test each model direction separately, keep embedding spaces isolated, and re-embed the full index. Prefer normalized schemas where practical, and retain source histories instead of relying only on model-written notes.

## Agent-Ready Context

Across **48 synthetic histories**, fixed-schema knowledge graphs barely changed after a writer-model swap, while compressed notes moved by **+9.91 or −13.28 percentage points** depending on migration direction. A half-migrated RAG index gained 4.96 points versus **11.90 points with full re-embedding**.

Treat model upgrades as memory migrations: test each model direction separately, keep embedding spaces isolated, and re-embed the full index. Prefer normalized schemas where practical, and retain source histories instead of relying only on model-written notes.

The evidence covers two open-weight models below 10B parameters and synthetic histories with exact answers. Store-only note repair missed the 90% recovery target in every case; raw history enabled recovery in **34 of 48 cases** for one tested direction, not universally.

## Connected Context

Feed7 judgment across 703 accumulated Signals:

This makes model portability a missing reliability constraint across the prior memory designs. It favors fixed, normalized, inspectable representations over model-authored compression when upgrades are expected, and shows that retrieval continuity requires full embedding migration rather than a mixed index. The limited models and synthetic histories narrow this to a concrete migration warning, not a universal ranking of memory architectures.

- [thedotmack/claude-mem](https://feed7.dev/p/claude-mem-1uba088) — Claude-Mem’s compressed persistent observations create the kind of model-written memory whose portability this study finds direction-dependent, strengthening the case for retaining source history alongside summaries.
- [UniMem: Complementary Episodic-to-Parametric Memory for Boundary-Agnostic Task Streams](https://feed7.dev/p/2607-26017v1-1opv1da) — UniMem’s learned consolidation raises an additional portability question: recurring knowledge moved into model parameters may be harder to preserve across model replacement than normalized episodic records.
- [Intelligence + Continual Learning = Expertise — Yu Su, NeoCognition](https://feed7.dev/p/intelligence-continual-learning-expertise-yu-su-neocognition-0f7dwfd) — The continual-learning vision depends on reusable state surviving upgrades; this study supplies a concrete stability requirement and shows that raw histories can sometimes recover knowledge that compressed notes alone cannot.
- [MemLens: A Value-Aware Memory Management System with Interactive Analytics for LLM-based Agents](https://feed7.dev/p/2607-25992v1-0q2kxfl) — MemLens treats retention as a quality and cost optimization, while this study adds migration robustness as another criterion for deciding which representations and source records to preserve.

## Context Map

- Layer: agent
- Domains: coding, data
- Topics: agent-memory, retrieval, agent-reliability

## Uncertainty

- The evidence covers two open-weight models below 10B parameters and synthetic histories with exact answers. Store-only note repair missed the 90% recovery target in every case; raw history enabled recovery in **34 of 48 cases** for one tested direction, not universally.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.