One memory under both agents, on your Mac. What Claude decided at 11 AM is what Codex knows at 2 PM. What either of them learned last Thursday is still there Monday. It survives compaction, survives the cap, survives switching machines, survives switching labs. And it proves itself in your first ten minutes or you were never charged.
One command. Card is not charged until it proves it remembers. Reply "goldfish" for a full refund, any time in 60 days.
>we keep stripe customer ids denormalized on the org row, the join was killing the dashboard.
noted. keeping the org row denormalized.
Stripe IDs denormalized on org row per your decision. Leaving as is.
Every session starts the same way. You re-explain the monorepo. You re-state the conventions. You re-identify the bug from yesterday. Auto-compact fires somewhere around 160K and quietly deletes the constraint from message 20. You catch it in the diff twenty minutes later. Claude hits the 5-hour window at 3:30, you move the work to Codex, and Codex has never heard of the thing you spent the whole morning on. So you type out the Stripe decision again. Third time this month. Codex flags it as a smell anyway.
Fifty to seventy-five percent of every session, spent re-learning what it already knew. You are the RAM.
"It's a goldfish."
"The user ends up serving as the continuity layer between sessions, which defeats the purpose of an AI coding assistant."
"Before compaction, Codex reported progress around 97%. After context was automatically compacted, Codex reported progress around 42%."
Anthropic shipped memory. It is machine-local and Claude-only.
OpenAI shipped memory. It is off by default, tied to a single codex_home, and Codex-only.
"Codex CLI cannot see Claude Code's context, and vice versa. They're built by different companies, store their working state separately, and have no channel between them."
That is not a bug on a roadmap. Anthropic is never going to teach Claude to remember what Codex did. OpenAI is never going to teach Codex to take notes for Claude. The most valuable memory in your entire workflow is the one that crosses the vendor line, and there is a structural reason nobody at either company is positioned to build it. Vendor lock-in, dressed as memory.
You run two labs because two labs beats one. You should not have to be the cable between them.
Tweety Bird hooks the lifecycle of both Claude Code and Codex. Every turn, on both sides, it sees what was asked, what was decided, and what was rejected, and it writes the parts worth keeping. Every turn, before either agent answers, it reads back the parts that are relevant right now. Deterministically. Not when the agent remembers to check. Not when a session goes idle and a 5,000-token summary gets extracted. Every turn.
Because it happens on the hooks, it does not matter which agent is working, which machine it is on, which plan you pay for, or whether the thread came from your phone or your terminal. The memory sits beside both agents, not inside either one. When Codex ships a new CLI, the memory does not notice.
And because the memory lives outside the context window, the context window can do whatever it wants to it. Nothing.
Under the hood it is MemoryRouter, which is in production and is the reason this is not another SQLite file of notes nobody queries.
pg pin alone."Memory entries are encrypted at rest and keyed to you. Team plan can pin memory to your own storage.
Auto-compact fires at 160K and summarizes what is in the window. The "why" was never in the window. 97 percent stays 97 percent.
Claude hits the window at 3:30. "codex, take it from here." Codex takes it, mid-thought. The cap is a handoff now, not a reset.
MacBook in the bag, mini on the shelf. Same memory on both. "How's the migration" gets the same answer no matter which box is doing the work.
Ask Codex to review what Claude wrote. It opens the PR already knowing the Thursday decision. It does not flag it. You did not type it a third time.
You believe memory has to prove itself in the first session or it is a note pile. So the install makes you find out.
Tell Claude Code one real decision about your repo. Something you have actually re-explained before.
Open Codex. Ask it about that decision.
If Codex knows, you are in. If Codex does not know, you close the terminal and you were never charged.
Your card is authorized at signup and captured only after the proof lands. Not "money back." Never charged.
Bring it. Bring the AGENTS.md and the DECISIONS.md you forget to update. Tweety Bird reads them on install and both agents carry what is in them from then on.
Then watch it shrink. Most people cut theirs to under 100 lines within a week, because the conventions live in memory now instead of a file that gets ignored past line 200. Keep the file for the things a new human needs. Let the memory carry the things the agents need.
| You want | What exists | Where it stops |
|---|---|---|
| Memory for Claude | Auto memory, claude-mem, Supermemory plugin | Machine-local, Claude-only, or crashy and token-hungry |
| Memory for Codex | Codex Memories | Off by default, one codex_home, Codex-only, not in the EEA or UK at launch |
| Memory across both | Mem0, Basic Memory, Supermemory via MCP | Wire into each tool yourself, metered retrievals, no phone, no schedule |
| One memory under both agents, written on every turn, on your Mac | Tweety Bird | Included in $29 a month |
Decisions and the why behind them. Rejected approaches. Conventions. Promises the agent made. Not your source. Memory entries are encrypted at rest and keyed to you. One command exports all of it as markdown. One command deletes all of it. Team plan can pin memory to your own storage.
The opposite. Re-explaining is where your tokens go: 94.8 percent of heavy-user tokens are cache reads, roughly 406 tokens read for every one written. Every decision Tweety Bird carries forward is 50 files you do not re-read at 9:30 AM. Recall is small and targeted, not a dump of everything it knows into every prompt.
Yes, because it was never in the context window. Auto-compact fires at 160K and summarizes what is in the window. The "why," the constraint from message 20, the rejected approach, all live in memory and get recalled when they are relevant again.
Yes. Memory that survives compaction and machines, plus a phone thread, files back, browser hands, and schedules, is already the product for one agent. The second agent is free when you want it. Most people who install for one end up running both, because the 5-hour cap stops being a wall once the other agent can pick up mid-thought.
Those are memory stores you wire into each tool yourself, hosted in someone else's cloud, with retrieval limits on the thing whose whole job is retrieval. No phone. No schedule. No hands. Tweety Bird's memory is one part of one install, written and read automatically on both agents, with a phone thread and a scheduler that use it.
Reply to any email with the word "goldfish." Full refund, 60 days, no form, no call. If it forgot something it should have remembered, we also want the screenshot, because that is a bug and founders' bugs ship first.
One command. Five minutes. Tell Claude one decision, ask Codex, and watch it answer. Your card is not touched until it does.
Install in one command$29 a month, locked for life for the first 500 machines. 417 left. Reply "goldfish" for a full refund, any time in 60 days.
Your laptop closes at 6. Your agents don't.
