codex - 💡(How to fix) Fix Topic-based memory directory with agent-initiated writes and /memory slash command [1 participants]

What feature would you like to see?

A topic-based memory architecture that combines Claude Code's memdir layout with oh-my-codex-style agent-initiated memory writes, plus a user-facing /memory slash command.

Problem Statement

1. Monolithic memory_summary.md does not scale: As memory grows, the single-file format becomes unwieldy for both the agent and users. It is hard to navigate, and loading the entire file into the system prompt wastes tokens.
2. Agent is forbidden from writing memories: The current developer instructions explicitly tell the agent "Never update memories. You can only read them.". This means when a user says "remember that I prefer X", the agent cannot persist it directly. Users must either wait for the async Phase-1/2 consolidation pipeline or manually edit memory files.
3. No CLI memory management: There is no slash command to list, edit, add, or clear memory topics without leaving the TUI.

Proposed Solution

1. Subfile Memory Directory (topics/*.md)

Introduce a topics/ subdirectory under ~/.codex/memories/ (and per-project memory roots) where each memory topic is an independent markdown file with YAML frontmatter:

~/.codex/memories/
  MEMORY.md              # Compact index / entrypoint
  memory_summary.md      # Legacy fallback
  topics/
    coding_style.md
    test_preferences.md
    api_patterns.md
    user_profile.md

Frontmatter schema:

---
name: Coding Style
description: Preferred patterns for this repo
type: project   # user | feedback | project | reference
keywords: [rust, clippy]
source: auto    # auto | user
updated_at: 2026-04-26T14:00:00Z
---

Benefits:

• Relevance-based loading: scan frontmatter, inject only the top-N most relevant topics (plus MEMORY.md index) into the system prompt, reducing token usage.
• Independent mutability: each topic can be read, written, or versioned independently.
• Freshness awareness: memories older than a threshold get a <system-reminder> caveat about potential staleness.

2. /memory Slash Command

Add a /memory (or /memories) command to the TUI:

This mirrors Claude Code's /memory edit UX and gives users explicit, immediate control.

3. Agent-Initiated Memory Writes (memory_write tool)

Expose a new built-in tool memory_write to the agent so it can persist knowledge in real time:

{
  "topic": "test_preferences",
  "content": "Always use pytest with the -x flag",
  "append": true,
  "type": "project",
  "keywords": ["pytest", "testing"]
}

Behavior:

• If append: true, append content to the existing topic body.
• If append: false (default), overwrite the topic file with new frontmatter + content.
• The tool writes only inside the memory root (~/.codex/memories/topics/), similar to how the consolidation sub-agent is sandboxed to the memory root.

Why this matters:

• Users can naturally say "remember that I prefer 2-space indentation" and the agent persists it immediately.
• Before context compaction, the agent can proactively checkpoint critical state.
• The existing Phase-1/2 pipeline continues to run in the background for automatic extraction, but explicit writes are instant and user-driven.

4. Scoped Memory (complementary to #18343)

The topic directory naturally supports scoping:

• Global memories: ~/.codex/memories/topics/
• Project memories: <project-memory-dir>/topics/
• Hybrid loading: load project topics first, then global topics, with project taking precedence.

Backward Compatibility

• If topics/ is empty, the prompt injection falls back to the existing memory_summary.md.
• Phase-1/2 consolidation can continue writing the old format while gradually migrating to topic files.
• No breaking changes to existing AGENTS.md or memory_summary.md behavior.

Related Issues

• #8368 — Long-term Memory (closed; this proposal extends the POC described there with a concrete directory structure and agent write capability)
• #18343 — Scoped memory management for Codex (complementary; topic directories make scoping trivial)
• #6532 — Add small list of commands to memory for common tool
• #3043 — Add memory functionality similar to Claude code

Are you interested in implementing this feature?

I have a working prototype of the memdir layout and /memory slash command in a local fork. I am happy to share a detailed design doc, help with implementation, or iterate on the approach if the team sees value in this direction.