1
00:00:00,740 --> 00:00:05,620
You know that feeling when you are three
weeks into building a complex feature with

2
00:00:05,660 --> 00:00:12,000
an AI agent, and suddenly it just, um,
completely forgets a core constraint you gave

3
00:00:12,040 --> 00:00:13,100
it on day one?

4
00:00:13,743 --> 00:00:14,403
Oh, totally.

5
00:00:14,823 --> 00:00:19,243
It gets buried under megabytes of subagent
chatter, terminal outputs,

6
00:00:19,603 --> 00:00:21,483
and intermediate reasoning steps, right?

7
00:00:22,103 --> 00:00:24,283
The human instruction just gets washed
out.

8
00:00:24,653 --> 00:00:25,393
Exactly.

9
00:00:25,633 --> 00:00:32,473
That exact problem is what OpenAI is
tackling with Codex CLI version 0.155.0.

10
00:00:32,993 --> 00:00:36,953
And hey, before we dive into the code,
quick shoutout to Jellypod for sponsoring the

11
00:00:37,013 --> 00:00:39,653
episode and helping us bring you this
daily breakdown.

12
00:00:40,200 --> 00:00:42,200
Yeah, huge thanks to Jellypod.

13
00:00:42,820 --> 00:00:48,960
So, Maya, this zero point one fifty five
point zero release introduces Memory V2,

14
00:00:49,600 --> 00:00:53,420
which is a total redesign of how Codex
handles long term state.

15
00:00:53,898 --> 00:00:57,138
Right, and it starts with complete storage
isolation.

16
00:00:57,638 --> 00:01:02,658
If you set memory version equals v2 in
your config file, or pass the environment

17
00:01:02,698 --> 00:01:08,498
variable CODEX MEMORY VERSION set to v2,
it spins up a completely separate database

18
00:01:08,558 --> 00:01:13,198
under your home directory in dot codex
slash memory slash v2 slash.

19
00:01:13,760 --> 00:01:14,400
Exactly.

20
00:01:14,920 --> 00:01:18,420
It leaves your legacy V1 context
completely untouched.

21
00:01:19,140 --> 00:01:22,420
But the really interesting part isn't just
where it saves things.

22
00:01:23,000 --> 00:01:26,900
It is the algorithmic shift in how
extraction actually works.

23
00:01:27,258 --> 00:01:28,158
Wait, how so?

24
00:01:28,542 --> 00:01:33,422
So, in V1, context chunking treated almost
all text equally.

25
00:01:33,462 --> 00:01:39,662
But in Memory V2, the extraction engine
explicitly enforces a rule to prioritize

26
00:01:39,715 --> 00:01:42,462
human evidence in memory v2 extraction.

27
00:01:42,900 --> 00:01:43,260
Ah!

28
00:01:44,060 --> 00:01:50,800
So explicit developer constraints,
feedback, and directives get weighted higher than

29
00:01:50,880 --> 00:01:54,140
thousands of lines of agent reasoning or
tool logs.

30
00:01:54,625 --> 00:01:55,665
Precisely.

31
00:01:55,729 --> 00:02:00,785
When the model condenses past sessions,
human input takes precedence during

32
00:02:00,838 --> 00:02:01,425
chunking.

33
00:02:01,612 --> 00:02:06,865
So your architectural rules do not get
overwritten by noisy subagent logs.

34
00:02:07,235 --> 00:02:10,015
That is huge for long running projects.

35
00:02:10,595 --> 00:02:15,115
But wait, if someone wants to test this,
do they have to wipe their existing memory

36
00:02:15,175 --> 00:02:16,975
setups or take a leap of faith?

37
00:02:17,873 --> 00:02:20,913
No, they built in a safe rollout
mechanism.

38
00:02:21,573 --> 00:02:24,993
You can set memory dual write equals true
in your settings.

39
00:02:25,817 --> 00:02:26,937
Dual writing mode?

40
00:02:27,333 --> 00:02:28,213
Yeah!

41
00:02:28,293 --> 00:02:33,653
It writes every memory update to both V1
and V2 storage engines in parallel.

42
00:02:33,797 --> 00:02:39,333
That lets you inspect readiness metrics
and verify accuracy before you cut over your

43
00:02:39,386 --> 00:02:40,693
workflow completely.

44
00:02:41,068 --> 00:02:45,968
Okay, but, um, running two extraction
pipelines in parallel...

45
00:02:46,608 --> 00:02:48,548
there has to be a cost trade off there,
right?

46
00:02:49,242 --> 00:02:50,242
There definitely is.

47
00:02:50,902 --> 00:02:55,482
Dual writing literally doubles your
background memory extraction API calls on every

48
00:02:55,522 --> 00:02:55,822
turn.

49
00:02:56,402 --> 00:03:00,462
So if you are on a tight token budget,
that is something to monitor closely.

50
00:03:00,820 --> 00:03:01,500
Makes sense.

51
00:03:02,120 --> 00:03:04,660
And what about summary only extraction?

52
00:03:05,042 --> 00:03:10,642
Well, V2 offers summary only extraction to
cut down prompt size,

53
00:03:10,682 --> 00:03:15,922
but if critical diagnostic evidence from a
tool run isn't explicitly anchored in

54
00:03:15,975 --> 00:03:21,202
human text, summary extraction might strip
out those raw diagnostic logs.

55
00:03:21,608 --> 00:03:27,228
Right, so if an error trace matters, you
need to reference it in your prompt so the

56
00:03:27,288 --> 00:03:30,748
extraction engine knows it is human
prioritized evidence.

57
00:03:31,208 --> 00:03:32,088
Spot on.

58
00:03:32,228 --> 00:03:37,928
Now, speaking of API calls and limits,
they also fixed a super annoying diagnostic

59
00:03:38,008 --> 00:03:38,968
bug in this release.

60
00:03:39,440 --> 00:03:40,700
Oh, what was that?

61
00:03:41,000 --> 00:03:47,160
In earlier builds, if your monthly billing
quota ran out, Codex returned an HTTP

62
00:03:47,400 --> 00:03:51,800
429 error that looked identical to a
temporary rate limit.

63
00:03:53,273 --> 00:03:54,093
Oh man!

64
00:03:54,573 --> 00:03:59,553
So developers were setting up retry loops
or waiting five minutes thinking it was

65
00:03:59,613 --> 00:04:04,713
just network traffic, when in reality
their credit card hit the monthly cap!

66
00:04:06,202 --> 00:04:06,862
Exactly!

67
00:04:07,402 --> 00:04:13,982
Now, 0.155.0 explicitly differentiates
quota exhaustion

68
00:04:14,442 --> 00:04:17,782
from standard concurrency rate limits in
error reports.

69
00:04:18,402 --> 00:04:20,302
No more chasing ghost retries.

70
00:04:20,877 --> 00:04:22,197
That is a relief.

71
00:04:22,737 --> 00:04:24,537
And what about the TUI changes?

72
00:04:24,857 --> 00:04:28,637
I heard the Agent Command Center got some
single key shortcut updates?

73
00:04:29,222 --> 00:04:29,542
Yes!

74
00:04:30,022 --> 00:04:34,302
In the agents overview, you can now press
h to hide background tasks,

75
00:04:34,922 --> 00:04:40,362
a to archive them, and d for confirmed
deletion of clean managed git worktrees.

76
00:04:40,737 --> 00:04:44,137
Confirmed deletion of clean managed
worktrees...

77
00:04:44,777 --> 00:04:48,357
so if you have background subagents
spinning up isolated worktrees,

78
00:04:48,797 --> 00:04:53,237
you can clean them up fast without
accidentally blowing away uncommitted code.

79
00:04:53,542 --> 00:04:54,182
Right.

80
00:04:54,235 --> 00:04:58,102
It checks that the worktree is clean
before allowing the deletion.

81
00:04:58,155 --> 00:05:02,102
It really tightens up managing multiple
background agent workspaces.

82
00:05:02,565 --> 00:05:04,545
Really sleek release overall.

83
00:05:04,985 --> 00:05:10,425
Memory V2 isolation, human evidence
priority, and cleaner terminal shortcuts.

84
00:05:10,905 --> 00:05:11,605
Good stuff.

85
00:05:12,543 --> 00:05:13,303
Absolutely.

86
00:05:13,883 --> 00:05:16,303
Alright, that is it for today's quick
breakdown.

87
00:05:16,763 --> 00:05:17,863
Catch you all next time!