Question 1

Codex already has memory. What does this add?

Accepted Answer

Codex has built-in Memories, including project summaries, durable entries, and supporting evidence from earlier chats. MemoryRouter adds one hosted vault that Codex can search and save to across sessions, repositories, and connected tools, so the decisions and rationale you discuss live outside the active context window. With the quick connect, Codex decides when to search and save. Add hooks and every request and final reply is captured automatically. Keep using native memory. The reason to add MemoryRouter is less reconstructing conversations and less managing what to save for next time.

Question 2

Why not use AGENTS.md, a handoff file, or a free plugin?

Accepted Answer

Those are valid options. AGENTS.md works well for stable instructions, and you can have your agent maintain a handoff file. Free memory plugins can automate parts of the workflow too. MemoryRouter is a managed alternative: saving and retrieval are connected to your coding sessions, and your supported tools share one vault. You pay for that ongoing service, not for an exclusive ability to store notes. Use the trial on real work and judge whether it reduces the explanations and memory maintenance you still do.

Question 3

Is it useful if I only use Codex?

Accepted Answer

Yes. Fresh-session continuity is the starting point. Codex saves decisions and reasoning to MemoryRouter and searches them in later sessions and after compaction. You do not need a second agent to use it. If you add Claude Code later, connect it to the same vault to share what you have already saved.

Question 4

What gets saved? What stays out?

Accepted Answer

With the quick connect, Codex saves what it decides is worth keeping, such as decisions and the reasons behind them. If you add hooks, they automatically capture your user requests and final assistant responses. Neither path saves your whole repository, and hooks do not capture tool inputs, tool results, terminal output, patches, or anything outside those visible messages. Decisions and reasoning mean what is written in those visible messages, not a model’s hidden reasoning. Text you put in a prompt or receive in a final reply can be saved, so do not include secrets there. MemoryRouter is a hosted service, not a local-only store. You can inspect and delete saved memories in your dashboard, and disable the integration for a repository with a .memoryrouterignore file.

Question 5

Will it bring the wrong project or an old mistake into my session?

Accepted Answer

MemoryRouter searches for context relevant to your current request; it does not load your entire history into every prompt. Codex and Claude Code use the same MemoryRouter vault across repositories, with repo tags providing context rather than separate memory silos. No memory system can guarantee perfect recall. Review and delete outdated memories in your dashboard, and keep your current instructions and repository state authoritative. Saved context survives compaction, but it does not guarantee task completion or resume unfinished work by itself.

Question 6

Will it slow Codex down or use more tokens?

Accepted Answer

Recall typically takes 300–500 ms. MemoryRouter retrieves relevant saved context rather than inserting your whole history. Recalled text does use space in the agent’s context and can add model input tokens. We do not promise a net token saving; the benefit to test is less repeated investigation and re-explanation. Your plan includes 200M memory storage and retrieval tokens each month, separate from your model’s inference usage.

Question 7

How do I connect it, and can my other tools share it?

Accepted Answer

Copy the setup prompt at the top of this page and paste it into Codex. It runs codex mcp add, you sign in to MemoryRouter once in your browser (you can create an account there), pick your vault, and approve. There is no API key and nothing to install. Open a new Codex session and it runs memory_status to confirm the connection. This path is model-directed: Codex decides when to search and save. For automatic capture on every turn, add hooks. That needs Node.js 18 or newer, Codex CLI 0.145 or newer, and a Memory Key, which stays out of chat. Other supported tools share saved memory when connected to the same vault. In browser-based ChatGPT and Claude, saving and recall are model-directed, not automatic capture of every chat. Connecting captures new conversations only; importing old chats is a separate workflow.

Question 8

What happens after the free trial?

Accepted Answer

The trial lasts 14 days. After that, the individual plan is $20 per month and includes 200M memory tokens each month across your connected tools. Additional usage is $0.50 per 1M memory tokens. These are memory storage and retrieval tokens, not model inference credits. MemoryRouter does not replace or include your Codex, OpenAI, or Claude subscription. Cancel anytime. Full plan details are on the pricing page.

Start a new Codex session. Pick up where you left off.

A fresh session
that already knows.

The context comes with you.

The expensive part
is explaining it again.

Keep the decisions.
Skip the repeat explanations.

Talk through the work.

Keep more than the conclusion.

Continue with the relevant context.

Keep AGENTS.md.
Lose the recap routine.

Codex’s built-in memory

AGENTS.md and handoff files

What MemoryRouter adds

Want every turn saved automatically?
Add hooks.

Don’t test “more memory.”
Test less repeating yourself.

Good questions.
Straight answers.

You already did the thinking.
Build on it.

Start a new Codex session. Pick up where you left off.

A fresh sessionthat already knows.

The context comes with you.

The expensive partis explaining it again.

Keep the decisions.Skip the repeat explanations.