AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Integrated Episodic and Semantic Memory via Modulating Transformer FeedForward Layers

Yiqun Yao, Xiang Li, Xin Jiang, Xuezhi Fang, Naitong Yu, Siwei Dong, Wenjia Ma, Jing Li, Aixin Sun, Yequan Wang

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add hypernetwork-based memory (HyperMem) over Generative AdapterHyperMem · Raw text retrieval (random access)F1: 0.297 -> 0.654Improved
use HyperMem instead of MemoryLLM-8B-1024 with similar memory sizeHyperMem · Raw text retrieval (from head)F1: 0.749 -> 0.876Improved
HyperMem vs MemoryLLM-8B-12800 which has far larger storageHyperMem · Raw text retrieval (from head)F1: 0.901 -> 0.876Mixed
HyperMem vs Generative Adapter on LongBenchHyperMem · LongBench HotpotQAF1: 0.385 -> 0.420Improved
HyperMem vs Generative Adapter on MultiFieldQA-ENHyperMem · LongBench MultiFieldQA-ENF1: 0.186 -> 0.213Improved
HyperMem vs full-context Llama3.1-8B on HotpotQAHyperMem · LongBench HotpotQAF1: 0.431 -> 0.420Hurt
HyperMem vs Generative Adapter on MSC personalizationHyperMem · MemGPT-MSCF1: 40.2 -> 43.0Improved
modulate parameters with memory injection on unrelated textHyperMem · FineWebEdu val perplexityperplexity: 7.66 -> 7.77No clear effect
modulate attention layers only or both attention and FFN instead of FFN only (ablation)HyperMem · Training loss (early)lossHurt
replace LoRA encoder with a randomly initialized shared Transformer layer (decouple layer) (ablation)HyperMem · Training loss (early)lossHurt
remove the hypernetwork and use memory as extended KV cache (no hyper) (ablation)HyperMem · Episodic memory loss (early training)lossNo clear effect
increase LoRA rank from 16 to 64 (ablation)HyperMem · Associative memory training lossloss: delta loss = 0.1Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy