Skip to content

Codex CLI Plugin

Semantic memory for Codex CLI. Shell hooks and a memory-recall skill, similar in architecture to the Claude Code plugin.

Why memsearch for Codex?

Codex CLI is OpenAI's open-source terminal coding agent. Unlike Claude Code (which has a mature plugin marketplace) or OpenClaw (which has a built-in memory system), Codex has no native memory plugin ecosystem. Hooks support is experimental, and there are few third-party memory solutions available.

memsearch fills this gap with a shell-hook-based plugin that gives Codex the same persistent memory capabilities as other platforms:

  • First-class memory for Codex -- no other solution provides hybrid semantic search with progressive disclosure
  • Same architecture as the Claude Code plugin -- if you're familiar with one, you understand both
  • Cross-platform portability -- memories captured in Codex are searchable from Claude Code, OpenClaw, or OpenCode
  • ONNX embedding default -- no OpenAI API key needed for the memory system itself (Codex uses OpenAI for the agent, but memsearch's embeddings are independent)

--yolo Mode and Sandbox

Codex CLI runs in a sandboxed environment by default. The memsearch plugin requires file system access to write memory files and run the memsearch CLI. The recommended approach:

  • Install option: The install.sh script configures hooks.json which works in any mode
  • Stop hook isolation: The Stop hook uses codex exec --ephemeral -s read-only with an isolated CODEX_HOME to prevent sandbox conflicts during summarization

If you experience issues with the Stop hook in strict sandbox mode, see Troubleshooting for diagnostic steps.

Key Features

  • Automatic capture -- conversations summarized via codex exec using gpt-5.1-codex-mini after each turn
  • Three-layer progressive recall -- search, expand, and drill into original rollouts (details)
  • Shell hook architecture -- similar to Claude Code plugin, easy to understand and modify
  • Orphan cleanup -- handles missing SessionEnd hook gracefully (Codex doesn't have one)
  • Milvus Lite lock handling -- automatically detects Milvus backend and skips concurrent index operations in Lite mode
  • ONNX embedding by default -- no API key required, runs locally on CPU
  • Local summarization fallback -- if codex exec fails, falls back to truncated raw text

When Is This Useful?

  • Codex as your daily driver. If you use Codex CLI for everyday coding, memsearch gives it memory that persists across sessions -- no more re-explaining context.
  • Codex + Claude Code workflows. Some developers use Codex for quick tasks and Claude Code for complex ones. memsearch provides unified memory across both.
  • Long debugging sessions. Codex sessions tend to be focused but context-heavy. memsearch captures the debugging trail so you can pick up where you left off.
  • Evaluating Codex. If you're comparing coding agents, having consistent memory across all of them provides a fair evaluation baseline.

Pages

  • Installation -- prerequisites, install, pre-cache, uninstall, updating
  • How It Works -- hook architecture, capture mechanism, memory files, Milvus Lite handling
  • Memory Recall -- three-layer progressive disclosure, comparison with Claude Code, manual invocation