Agent Memory
Evolving, versioned and multi-representation memory infrastructure.
I'm Wang Zupeng, a master's student exploring the infrastructure behind intelligent systems, from Agent memory and KV cache to distributed storage.
Research premise
Memory is not a static record. It is a living system of representations, versions, and decisions.
Evolving, versioned and multi-representation memory infrastructure.
Reuse, invalidation and lifecycle management for LLM inference.
Distributed storage, NFS, Ceph and data movement across tiers.
Reliable data services and performance-oriented engineering.
Featured research
2025 — PresentVersioned multi-representation storage
This work studies how text, embeddings, summaries, and execution-specific KV-cache views can remain version-consistent while being scheduled across GPU, CPU, and SSD resources.
“KV Cache is not a factual memory source, but a model-execution-specific materialized view.”Read the research direction
Selected work
View all projectsBackend components for strategy trading backtesting, including system and professional backtesting, community indicators, and short-term strategy data processing.
An investigation into cache hit-rate observability, request paths and measurement boundaries in NFS-Ganesha deployments.
A compact static delivery stack built on Astro, Azure, Nginx and Cloudflare, with reproducible deployment from source.
Field notes
Browse the notebook