mnemo

Codex memory surfaces

Native model

Codex does not expose a separate local semantic memory store comparable to mnemo. Its durable, user-controlled memory-like surface is instruction aggregation from AGENTS.md / AGENTS.override.md files plus configured fallback filenames.

OpenAI’s agent-loop documentation describes Codex adding user instructions from $CODEX_HOME and then from folders between the Git/project root and the current working directory, subject to a default 32 KiB project-doc budget. The Codex source describes the same discovery path and concatenation model.

Locations

Global/user scope:

• $CODEX_HOME/AGENTS.override.md, otherwise $CODEX_HOME/AGENTS.md.
• $CODEX_HOME is normally ~/.codex.

Project scope:

• For each directory from project root to current working directory: AGENTS.override.md, then AGENTS.md, then configured project_doc_fallback_filenames.
• The default project root marker is .git when no custom project_root_markers are configured.

Shape and retrieval

• Plain markdown instruction files.
• Files are concatenated from broad to narrow scope, so more specific files appear later.
• Empty files are skipped.
• Project docs are bounded by project_doc_max_bytes.
• Codex does not walk past the project root for project docs.
• There is no documented MEMORY.md-style index: hierarchy affects which instruction files are loaded, not where separate memory bodies are stored.

Conversion notes for mnemo

Recommended import order:

1. Repo root AGENTS.md and nested project AGENTS.md files.
2. AGENTS.override.md only with explicit opt-in because it may be personal or temporary.
3. $CODEX_HOME files only with explicit user-scope import.
4. Configured fallback filenames only when the importer can read Codex configuration safely.

Suggested mapping:

• Build/test commands -> config.
• Architecture and workflow conventions -> pattern.
• Explicit decisions or rationales -> decision.
• Known gotchas -> discovery.

Important importer behavior:

• Preserve directory scope so nested instructions do not become global project memory accidentally.
• Record whether a source was an override file.
• Do not infer a native Codex memory database; treat AGENTS.md as prompt-level durable context.

Sources

• https://openai.com/index/unrolling-the-codex-agent-loop/
• https://raw.githubusercontent.com/openai/codex/main/codex-rs/core/src/agents_md.rs

This site is open source. Improve this page.