Kevin Lin

Kevin Lin 林冠言 builds agentic systems at Letta. He coauthored sleep-time compute, which moves reasoning about available context ahead of a user's question, and MemGPT, which manages memory beyond a language model's limited context window. His personal profile describes his work at Letta; Berkeley lists him as a Member of Technical Staff there.[1][2][3][4]

Lin coauthored the April 17, 2025 paper Sleep-time Compute: Beyond Inference Scaling at Test-time with Charlie Snell, Yu Wang, Charles Packer, Sarah Wooders, Ion Stoica and Joseph E. Gonzalez. The team addressed a recurring cost in stateful applications: related questions about a document, repository or conversation can require the model to repeat the same background reasoning.[3]

Their method gives the model context before it receives the question. The model draws inferences, anticipates possible questions and rewrites the context into a natural-language representation useful for later answers. A coding assistant, for example, could identify architectural patterns or potential debugging strategies while idle. The resulting context can be reused across related queries, spreading the preparation cost across several answers. The implementation uses a rethink_memory function to replace the context and a separate function to end preparation.[3]

The coauthors reported approximately fivefold less test-time compute for comparable accuracy on Stateful GSM-Symbolic and Stateful AIME, modified math benchmarks that separate background context from the question. This shifts work into preparation rather than establishing a fivefold reduction in total compute. Gains varied by model and budget: o1 showed limited gains, and some higher-budget settings favored reasoning only after the question arrived. On Multi-Query GSM-Symbolic, their modeled cost analysis reported up to a 2.5-fold decrease in average cost per query relative to the single-query baseline, with ten questions sharing a context and assuming test-time tokens cost ten times as much as preparation tokens.[3]

The team also tested advance repository exploration on SWE-Features, a benchmark of 33 feature-oriented pull requests from Aider and ComfyUI. They reported roughly a 1.5-fold reduction in test-time tokens at lower budgets, but evaluated overlap between predicted and actual modified-file sets rather than functional correctness. Higher budgets could favor the test-time-only baseline. The coauthors identified query predictability as a central condition: advance reasoning is less useful when the eventual question is difficult to anticipate or unrelated to the available context.[3]

Before the sleep-time compute collaboration, Lin coauthored MemGPT: Towards LLMs as Operating Systems, first submitted on October 12, 2023 and revised on February 12, 2024. His collaborators were Charles Packer, Sarah Wooders, Vivian Fang, Shishir G. Patil, Ion Stoica and Joseph E. Gonzalez. This earlier work supplies context for his research on persistent information: MemGPT uses hierarchical memory tiers and data movement to extend usable context, with interrupts controlling interaction between the system and user. The team evaluated it on document analysis and multi-session conversations.[4]

Lin completed his PhD in EECS at UC Berkeley, where he worked with Dan Klein and Joey Gonzalez. The Berkeley NLP directory lists him as a 2024 PhD alumnus; his subsequent work at Letta includes the 2025 sleep-time compute collaboration.[1][2][3]