🛰️ Daily AI Frontier
‹ back to 2026-09-07

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

Research LLM Agents

Ranking

Overall 81
Content 100
Popularity 37

Observed public metrics from 1 member.

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

Merged summary

TL;DR - This controlled study finds that an AI agent’s stored memory may not transfer reliably when its underlying model changes. Fixed-schema knowledge graphs were highly portable, while model-generated notes and partially migrated RAG indexes suffered substantial degradation.

  • KG-fixed accuracy changed by just (+0.0004 \pm 0.0020) after a writer-model swap.
  • NOTES were strongly model-coupled, with migration changing accuracy by (+9.91) or (-13.28) percentage points depending on direction; 80% of their deficit came from information lost during construction.
  • A 50/50 mixed-embedding RAG index gained only 4.96 points versus 11.90 points from full re-embedding, with retrieval failures accounting for 81% of the RAG deficit.
  • Store-only NOTES repair never achieved 90% recovery across 48 histories, while retaining raw histories enabled recovery in 34 cases for one migration direction.

Sources (1)

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

arXiv cs.AI Ankit Goyal, Jaideep Ray 2026-09-04 arXiv:2609.05339
Public signals Semantic Scholar citations 0 · Semantic Scholar influential citations 0
Providers: Hugging Face · N/A OpenAlex · N/A Publisher · N/A Semantic Scholar · Citations 0 · Influential citations 0 X · N/A Fetched 2026-09-13 14:08:13.939297 UTC

TL;DR - This controlled study finds that an AI agent’s stored memory may not transfer reliably when its underlying model changes. Fixed-schema knowledge graphs were highly portable, while model-generated notes and partially migrated RAG indexes suffered substantial degradation.

  • KG-fixed accuracy changed by just (+0.0004 \pm 0.0020) after a writer-model swap.
  • NOTES were strongly model-coupled, with migration changing accuracy by (+9.91) or (-13.28) percentage points depending on direction; 80% of their deficit came from information lost during construction.
  • A 50/50 mixed-embedding RAG index gained only 4.96 points versus 11.90 points from full re-embedding, with retrieval failures accounting for 81% of the RAG deficit.
  • Store-only NOTES repair never achieved 90% recovery across 48 histories, while retaining raw histories enabled recovery in 34 cases for one migration direction.
item →