lemory

Lemory is a local-first context database middleware for AI agents, integrating existing markdown files (e.g., Obsidian vaults) into a system of memories, knowledge, and skills. It uses actual files as the database without virtual file systems or hosted DBs, allowing for zero-cost migration and compatibility with standard tools like `ls`, `grep`, and `git`. Core features include a 4-way hybrid fusion search combining semantic vector search, Korean-optimized keyword search (BM25), wiki-link graph traversal, and time-window search. The system is specifically tuned for Korean and CJK languages, offering high search accuracy through typo correction, inflectional stem matching, and bigram indexing. A 3-tier loading structure (L0 summary, L1 outline, L2 raw) enables significant token savings by allowing the AI to decide reading depth at runtime. For agent working memory, Lemory provides tools like `remember`, `recall`, `reflect`, and `resume_case`, supporting a 'memory pyramid' where conversation content is distilled into atomic facts and promoted to scene narratives and persona notes. It operates as an MCP server, enabling integration with Claude Code, Cursor, Windsurf, and other AI coding agents, while also offering an OpenAI-compatible API proxy for non-MCP clients. All queries and writes are logged locally in SQLite for timeline observation, and AI-generated notes are managed with source attribution and rollback buttons.