kevintsai1202/deep-memory

Deep-Memory is a skill for AI coding assistants, such as Claude Code or Codex, that gives them long-term memory across separate conversations. Normally an AI assistant forgets everything once a session ends. This project runs in the background, automatically pulling up relevant past experience during a conversation, and once a task is finished successfully, it writes the approach that worked into a private knowledge base so it can be reused later. The README describes it as a restructured and renamed version of an earlier project called auto-skill, rebuilt to separate code from stored data and to add local hybrid search. The README describes a five step loop that runs on every conversation turn: extracting keywords to fingerprint the current topic, detecting whether the topic has changed and the knowledge base needs rereading, checking for skill-specific pitfall notes if a particular tool was used, retrieving matching general knowledge base entries, and finally writing back a summary once a task appears complete. It also describes cross-skill memory, where invoking one tool can trigger a reminder based on something learned during a past use of that same tool, and proactive capture, where the assistant asks whether to save a solution once a problem seems solved rather than requiring the user to take notes manually. Technically, retrieval combines semantic search with keyword based BM25 search through a local ChromaDB database, re-ranked using a model called BGE-Reranker-base that runs on the CPU. Knowledge is split into a hot store of curated, indexed entries and a cold store of raw, unrefined notes that get promoted into the hot store once enough accumulate. A separate backup component can push the knowledge base to a private GitHub repository and restore it on another machine, checking first whether the remote copy is newer so devices do not silently overwrite each other. Installation can be done three ways: copying the skill files into an AI agent's global skill folder so every project shares one knowledge store, using a quick install command that fetches the package from GitHub directly into that global folder, or copying the skill folder into a single project for a self-contained setup. By default all of a user's knowledge is stored under a folder in their home directory shared across projects, though this can be changed to keep separate stores per project.