Show HN: ChatIndex – an open-sourced long-context managment system (github.com)

ChatIndex is an open-source context-management system that builds hierarchical, dynamic-resolution indexes over long conversations so LLMs can retrieve relevant parts of a dialogue without losing raw data. Aiming to solve “context rot” and the limits of lossy memory summaries, it preserves full conversation history and adds multi-resolution access: high-level topic nodes are returned when sufficient, while deeper message nodes are fetched only when needed. This makes a single coherent conversation thread feasible for assistant apps instead of juggling multiple short chats. Technically, ChatIndex adapts PageIndex for conversational data by incrementally generating a Context Tree (CTree) via LLM-driven topic-switch detection. The CTree is temporally ordered—new topics can branch only from the current node or its ancestors—and contains TopicNodes (topic_name, summary, start/end indices, children) and MessageNodes (system/user/assistant messages, message_index). Controls like max_children limit layer width to bound retrieval cost, and offline post-processing can optimize tree structure. The repo includes a demo for building CTrees from histories and uses the OpenAI API for topic detection. For AI/ML practitioners, ChatIndex offers a practical, transparent alternative to lossy memory stores, enabling more scalable, explainable long-context retrieval and easier integration into LLM apps.