What we measured, and what we have not
Everything below comes from our own tests on one Windows computer with real Codex chats. We publish the method and the limits so you can judge them.
How each test works
We plant a codeword, an open to-do, a decision (one of six options) and eight unguessable six-digit numbers, each produced by a tool so the model cannot guess them. They are spread across about a dozen filler paragraphs and six compactions (run with /compact). At the end we ask for all eleven. A script reads the chat record and scores every answer. Reasoning effort is Light.
GPT-5.6 Sol
| Test | Result |
|---|---|
| Six compactions, plugin off (one run) | 0 of 11 recalled |
| Six compactions, plugin on (four runs, including the finished product installed the way customers install it) | 11 of 11 recalled each time |
| Heavier test: 30 separate findings, six compactions | 33 of 33 |
| Codex's own automatic compaction (20 findings) | 22 of 23. The one miss was the model mis-copying a number; the saved summary held the true value. |
On Sol, Codex's own compaction kept only its standing instructions and the last few messages. Everything older was gone.
GPT-5.6 Terra and GPT-5.6 Luna
| Model | Result |
|---|---|
| GPT-5.6 Terra, plugin off | 11 of 11 |
| GPT-5.6 Terra, plugin on | 11 of 11 |
| GPT-5.6 Luna | Codex already kept most of the conversation on its own in earlier tests of a previous version. |
On these models Codex keeps all of your messages plus its own long summary, so there is little for Seamless Context to add. We checked Terra at High effort as well: the same long summary was present.
How the summary holds up as a thread grows
In offline stress tests with realistic filler, every labelled fact was kept up to about 100 separate findings (96 percent when the facts were stated in running prose), and decisions and instructions were kept at every size we tried. Beyond that the oldest, least specific facts are dropped first: at 150 findings about 85 percent of labelled facts were kept, at 300 about 68 percent.