zhexulong/context-manager

The context-manager project turns agent conversations into reusable guidance and context. The README is mostly in Chinese with one English subtitle. The premise is that after an AI coding agent has been used for a while, the chat history grows long but the genuinely useful conclusions get harder to find: either important context gets buried, or it gets dropped when the session is compacted or ends. This tool tries to extract the parts worth keeping so that later agent sessions can pick them up directly. The core flow starts from a conversation, runs a refining step, writes the result into a working-memory folder, refines again into a guidance folder, and then later agents read from guidance on demand. The guidance area is split into three plain categories: decisions, processes, and conventions. The README states the tool currently targets the Codex agent and that support for other agents such as Claude Code is planned next, not built yet. Two install paths are documented. The npm path is to install the @omnimer/context-manager package globally and then run context-manager init --repo-root . inside the target repository. The source path clones this repo, runs npm install, and then runs the bin script with --repo-root pointing at the target. The init step writes a Codex hook configuration into the target repo that points at the already-installed runtime, instead of copying the hooks or scripts folders. A short list of common commands names npm test, a compile script, a lint script, a replay-pending script, and a smoke end-to-end script. The current status section says the integration goes through Codex hooks, that working-memory and guidance are produced, and that a guidance-recall mechanism lets later sessions pull what they need. Other notes describe default prompt templates under scripts/prompts that can be overridden through an environment variable, report folders for pending replays and LLM failure logs, project-level hooks declared in .codex/config.toml, and a guidance-recall skill installed under the user's home Codex skills folder. The compile step takes the latest verified state as authoritative when sources conflict, so stale guidance is not preserved. The README ends with a list of reference repositories that inspired the design, including claude-memory-compiler, openai's codex and codex-plugin-cc, and zilliztech's memsearch, and an invitation to file issues and pull requests, with thanks to the linux.do community.