Hermeneutic 将 AI 纠正记录转化为后续任务的上下文

You ask a coding agent to review a function. It starts refactoring. You steer it back to the review you requested. That exchange contains useful information about your intended scope. Hermeneutic, built by Hermes Labs, mines corrections from agent conversations and surfaces relevant guidance when similar tasks come up. In this example, a later code-review request could receive guidance about staying within the requested scope—drawn from your earlier corrections. Your conversation history becomes a source of context for new work. How correction memory works Hermeneutic reads supported Claude Code, Codex, and OpenAI-format logs. It recognizes correction-shaped user turns and records the surrounding exchange when available: your request → agent response → your correction → repaired response Enter fullscreen mode Exit fullscreen mode The resulting local corpus preserves the episode around each correction. The bucket command lets you inspect recurring categories, including missed constraints, wrong targets, unnecessary confirmation, and scope expansion. Retrieval makes those records useful during another task. Using Ollama embeddings, Hermeneutic matches a new prompt against earlier prompts in the corpus. Relevant matches supply correction categories, which become a compact guidance preamble through category templates. Each advice bullet cites its supporting correction records. For a release task, that might surface your prior guidance to cite command output when reporting completion. For a similar review task, it might surface guidance about respecting the requested scope. Put your own history to work With Python 3.10+ and Ollama available: python -m pip install hermeneutic==0.1.12 hermeneutic mine ~/.claude/projects \ --format claude-code \ --glob '**/*.jsonl' \ --out ~/.hermeneutic/triples.jsonl hermeneutic bucket ~/.hermeneutic/triples.jsonl ollama pull nomic-embed-text hermeneutic compile-index --triples ~/.hermeneutic/triples.jsonl hermeneutic compile 'Review this function and report your findings.' Enter fullscreen mode Exit fullscreen mode When matching corrections meet the relevance threshold, the final command emits guidance for that prompt. Use it directly in your workflow, or install the optional Claude Code hook to supply retrieved guidance as prompt context: hermeneutic install-compile-hook Enter fullscreen mode Exit fullscreen mode Start with the corrections you already made. Try Hermeneutic on GitHub, follow the memory setup, and star the repository to follow its development. What correction do you keep repeating to your coding agent? Built by Hermes Labs. Prepared by our agent infrastructure from the project README and documentation.