WEBVTT

1
00:00:00.550 --> 00:00:02.278
AI tools are getting smarter.

2
00:00:03.250 --> 00:00:05.255
But their memory is still fragmented.

3
00:00:06.450 --> 00:00:08.509
Claude does the work. Codex starts tomorrow. And

4
00:00:08.509 --> 00:00:10.567
the context disappears.

5
00:00:12.650 --> 00:00:13.813
So I stopped treating the model as the

6
00:00:13.813 --> 00:00:14.975
system.

7
00:00:16.150 --> 00:00:17.153
I built Loom.

8
00:00:18.000 --> 00:00:20.389
One persistent memory for every project.

9
00:00:21.550 --> 00:00:23.178
Then I built zai. One command that routes

10
00:00:23.178 --> 00:00:24.807
the work to Codex, Claude, or a local

11
00:00:24.807 --> 00:00:26.435
model.

12
00:00:27.850 --> 00:00:30.282
If one provider fails, another takes over.

13
00:00:31.450 --> 00:00:33.818
The worker can change. The memory stays.

14
00:00:35.200 --> 00:00:37.291
Every decision. Every failure. Every test. Every commit.

15
00:00:37.291 --> 00:00:39.381
Every next step.

16
00:00:40.650 --> 00:00:43.039
Claude knows what Codex did. Codex knows what

17
00:00:43.039 --> 00:00:45.429
Claude did. No manual handoff.

18
00:00:47.000 --> 00:00:49.795
GitHub stores the code. Loom remembers why.

19
00:00:51.200 --> 00:00:53.419
Now I open a terminal, type one command,

20
00:00:53.419 --> 00:00:55.637
choose the project, and keep building.

21
00:00:57.550 --> 00:00:59.491
Inside a tracked project, zai skips the picker

22
00:00:59.491 --> 00:01:01.433
and gets straight to work.

23
00:01:02.600 --> 00:01:04.637
The memory and orchestration stay local-first on my

24
00:01:04.637 --> 00:01:06.675
Fedora workstation.

25
00:01:07.350 --> 00:01:09.558
Codex and Claude can remain external. Local models

26
00:01:09.558 --> 00:01:11.766
are optional workers.

27
00:01:12.350 --> 00:01:14.889
One project. One memory. Multiple minds.