SimpleMem: Efficient Lifelong Memory for LLM Agents
Notes on Liu et al. (UNC Chapel Hill) — a three-stage memory pipeline that trades write-time compression for read-time efficiency, with multi-view indexing and adaptive retrieval planning.
Research Briefs
Sub-Millisecond Memory: How a 4-Slot Register Solves Multi-Turn Conversations
How a tiny 4-slot context register solves multi-turn conversation tracking for on-device personal assistants — under a kilobyte, under a millisecond, no cloud required.
Miras: It's All Connected — Unifying Sequence Models Through Memory Objectives
Notes on Behrouz et al. (Google Research) — how Transformers, Mamba, Titans, and friends are all just associative memories with different loss functions, and what happens when you explore beyond the two objectives everyone uses.
Memory in Large Language Models: Mechanisms, Evaluation and Evolution
Notes on Zhang et al.'s survey (arXiv 2509.18868) — how to think about and evaluate LLM memory as a system property, not just a bag of techniques.
Cognitive Memory in Large Language Models — A Distributed Systems Perspective
Notes on a survey from Li Auto — and why the LLM memory problem looks a lot like distributed systems problems I've worked on before.