ramennuts/obsidian-engram

Engram is a memory system for Claude Code, the AI coding agent, that helps it remember context between separate work sessions instead of starting fresh and forgetting everything each time. It is built entirely on plain markdown files stored in an Obsidian vault, a popular note taking app, so there is no database, no paid service, and no network calls involved. Everything can be opened and edited as an ordinary text file. The system splits memory into two layers based on how quickly things change. The first layer, durable memory, holds facts that stay true across all sessions, such as who the user is, their preferences, and background facts about the project. This layer loads at the start of every session. The second layer, working memory, tracks the current state of active work: a live status board, a queue of tasks that are active or blocked, and dated handoff notes written at the end of each session summarizing what was done and what is still open. Two small automated pieces tie this together. A hook that runs automatically when a session starts injects a short summary of the current state, so the agent already knows the context without being told to go read anything. A companion skill lets you write a detailed handoff note with one command at the end of a session, which then gets picked up automatically next time. A separate weekly script archives finished tasks out of the active queue so the injected summary stays short and useful as the vault grows over time. The toolkit also includes a script that checks the whole setup is wired together correctly, and another that checks the durable memory files for broken links or bad formatting. Installation takes about five minutes using a single setup script that seeds the two memory folders and installs the necessary pieces. The project is released under the MIT license.