This repository is the official code release for the Memento research project by Microsoft AI Frontiers. Memento introduces a method for large language models to learn context management by segmenting chain-of-thought reasoning into blocks, compressing each block into a concise 'memento' (summary), and then evicting the original block from the KV cache during generation. This achieves significant peak KV cache reduction (2-3x) and throughput improvements (1.75x) with minimal accuracy loss, which can be recovered via reinforcement learning (CISPO). The repository is a hybrid of a data research pipeline and a software application overlay. The data pipeline (data/) ingests raw chain-of-thought traces from the OpenThoughts dataset, filters them by domain (math, code, science), and transforms them into the Memento format. This involves LLM-based boundary scoring, dynamic programming segmentation, and iterative LLM-based summarization with judge feedback, resulting in the OpenMementos dataset (228K traces). The application component is a surgical overlay on the vLLM inference engine (vllm/). It adds a BlockMaskingProcessor that acts as a token-by-token state machine, detecting block boundaries and orchestrating physical KV cache compaction. It also includes a BlockLengthCapLogitsProcessor to enforce block size limits. This overlay is designed to be patched onto a standard vLLM 0.13.x installation, modifying only the necessary Python files while restoring compiled .so interfaces for compatibility. Overall, this is a high-quality research codebase demonstrating strong engineering craftsmanship. The data pipeline is robust and well-structured, and the vLLM overlay is surgically precise. However, it has a high technical drift risk due to its tight coupling to a specific vLLM version and its reliance on proprietary LLM APIs (GPT-5.x) for the data generation pipeline, which limits full reproducibility without access to those specific models.