Phudish Prateepamornkul
AI Scientist at SCB DataX
About Me
Hi! My name is Tam and I am an AI Scientist at SCB DataX, and a graduate researcher at Berkeley AI Research, where I work with Akshat Gupta in the Berkeley Speech Group.
I work on memory in language models: how a model holds on to what it knows, how that store gets updated as new information arrives, and what breaks when it has to carry a lifetime of context instead of a single prompt. I have come at this from three levels of the stack — editing the weights directly (knowledge editing), compressing long context into the KV cache (cartridges), and building external memory stores for agents (extraction, retrieval, agentic navigation). The same question shows up at every level: what do you keep, what do you throw away, and what happens when the model has to change its mind?
Lately I am most interested in KV-cache memory — training small KV caches that stand in for a long corpus, and making them continual, so that gating and routing over sub-caches lets new documents be written in without erasing what is already there. Alongside that I build controlled evaluation for memory systems at the 1M–10M token scale, because I do not think we yet know which of these paradigms degrades gracefully and which one just looks good on short benchmarks.
Research Interests
- KV-Cache & Long-Context Memory — context compression, cartridges, test-time training of compact memory representations
- Continual Learning for Memory — gating and routed sub-cartridges so memory can be updated without catastrophic forgetting
- Knowledge Editing & Interpretability — lifelong editing at scale [EMNLP'25 Findings]
- Evaluating Agent Memory — comparing extraction, retrieval, long-context, and agentic navigation under matched cost and judging