LinkedIn Paper Tests Memory Portability Across Models: Notes Swing ±10-13 Points, Knowledge Graphs Barely Move
dair_ai · x · 2026-09-07
A new LinkedIn paper tackles an overlooked question: can an agent's memory survive swapping the underlying model? The team stored the same agent history four ways—verbatim long context, chunked for retrieval, model-written notes, and a fixed-schema knowledge graph—then swapped the model doing the reading.
Results diverged sharply:
- Fixed-schema knowledge graphs barely moved: just 0.0004 accuracy change across a writer swap
- Model-written notes were volatile: accuracy shifted +9.91 or -13.28 points depending on swap direction
dair-ai calls memory one of the hardest things to get right when building agents and worth deliberate optimization. As users increasingly move between models, agents, and providers, memory portability is becoming a core engineering concern—especially for anyone who upgraded an agent's model only to watch it forget.
More from coding & agent
- Shopify design lead on AI-era workflow: designers are becoming full-stack editors — KlausCodes · 2026-10-26
- Ultrametric open-sources Open Startup Repo to let AI agents run startup operations — garrytan · 2026-10-26
- MCP fault injection study: agent task scores drop 60% under realistic tool failures, scale doesn't help — dair_ai · 2026-10-26
- General agents burn ~140M tokens/month on LinkedIn outreach; scripts do it 60x cheaper — ChrisHarpon2 · 2026-10-26
- OpenAI upgrades Decisions API: 400 questions per call, native voice and image support — OpenAIDevs · 2026-10-26
- NLI alone rejects 5/6 new agent memories; a 0.45 similarity gate fixes it all — Alternative_Pin9598 · 2026-10-26