What is paperguru-benchmark?

PaperGuru is a research project introducing a new memory architecture for AI agents that work on long, complex tasks. The core problem it addresses is that AI language models are good at answering questions within a single conversation, but they struggle to keep track of information across very long sessions, like spending multiple days reviewing research papers or reproducing a published study from scratch. The solution is called Lifecycle-Aware Memory, or LAM. The idea is that memory for AI should not just store facts but also track when facts become outdated, how documents relate to each other through citation chains, and where each claim originally came from. The system's central mechanism, called Capital Chunk Memory, separates memory into a compact routing layer and the full raw content, allowing the agent to find relevant information efficiently even as its knowledge archive grows without limit. The repository contains the benchmark results, paper, and evaluation materials demonstrating the system's performance. On two published research benchmarks, PaperBench and SurveyBench, PaperGuru outperforms the previously best-known results by a substantial margin. This is primarily a research artifact you would use if you are studying AI agent memory systems or evaluating long-horizon AI agents on scientific research tasks. It is written in TeX, reflecting its academic paper orientation.