codex-memory

English | 简体中文

Curated long-term memory for Codex CLI, delivered as an MCP server. It stores atomic facts as plain Markdown files, deduplicates on write, retrieves via a lightweight index, and — crucially — forgets: entries decay by type-aware half-life and are archived after human review.

Core idea: raw experience accumulates, but knowledge is what survives distillation — so forgetting is treated as a feature on equal footing with remembering.

• 4 MCP tools: memory_index / memory_search / memory_write / memory_revise
• One memory per Markdown file, git-versioned, human-auditable — no database
• Built-in dedup (weighted token overlap, CJK-aware), source ranking (user-correction wins)
• Global/project scopes prevent repository-specific knowledge from leaking into unrelated tasks; optional provenance points back to supporting sessions
• Development schema is strict: cwd, scope, counters, dates, type, and source must satisfy the current contract; no legacy defaults or migrations
• Decay scoring, an archive buffer with 90-day expiry, driven by a small CLI
• Explicit memory graduation promotes proven global preferences/lessons into a managed AGENTS.md region and retains a permanent audit copy
• Weekly consolidate job — semantic dedup, pattern promotion, and an episodic bridge from Codex's built-in memories, powered by codex exec (no extra LLM dependency), one git commit per run
• Built-in evals: repeat-mistake rate, write-burst and project-coverage signals (eval), plus the index growth curve (growth)

Documentation

Agent instruction templates (append to ~/.codex/AGENTS.md): English · 中文

License

MIT © Xiandong Liu (刘宪东)