XNews

Claude-Reflect: Plugin to Fix Agent Memory Across Sessions

Claude-Reflect: Plugin to Fix Agent Memory Across Sessions

Claude-Reflect is a Claude Code plugin that captures corrections and preferences during sessions and, upon running /reflect, presents them for approval before writing into CLAUDE.md files. It supports multiple config files (CLAUDE.md, AGENTS.md) and works with Codex and Cursor. v2 adds /reflect-skills to suggest new commands from repeated patterns.

Key insights

Capturing and Persisting Agent Corrections

The plugin hooks into session interactions: when a user corrects the agent, praises it, or says 'remember:', the hook captures the context and queues it. Running /reflect presents a table of items for the user to approve, modify, or skip. Approved items are written into CLAUDE.md, ensuring future sessions inherit the knowledge. It supports multiple config files including project-level and subdirectory CLAUDE.md and AGENTS.md.

Skill Discovery from Repeated Patterns

Version 2 adds /reflect-skills, which scans past two weeks of session history to identify repeated questions or commands. For example, if the user asked 'check my efficiency today' many times, it suggests creating a command. The plugin uses AI semantic filtering to handle corrections in any language, not just English keywords. Each suggestion includes a confidence score and requires human approval before writing.

Source posts