---
title: "Codex memory v2, with and without past.dev"
description: "We tested the new Codex memory, v2, on GPT-6.1 Sol. It passed 66.3% of 600 tasks against 49.7% for v1. With past.dev on top, Codex passed 85.8%."
canonical: https://past.dev/blog/codex-memory-v2-past-dev
date: 2026-11-05
category: Announcements
authors: The past.dev team
---
# Codex memory v2, with and without past.dev

OpenAI just gave Codex a new memory, "v2". We ran 600 tasks on GPT-6.1 Sol three times: with memory v1, with memory v2, and with memory v2 and the past.dev plugin.

![Codex on GPT-6.1 Sol. Tasks passed: 49.7% with memory v1, 66.3% with memory v2, 85.8% with memory v2 and past.dev. Input tokens per task: 154k, 184k, 109k. Cost per passed task: $0.138, $0.119, $0.055.](/blog/codex-memory-v2/results.svg)

Memory v2 is an upgrade: Codex passed 66.3% of the tasks with it, against 49.7% with v1. With the past.dev plugin added to v2, Codex passed 85.8%, read 41% fewer input tokens per task, each passed task cost 54% less, and the median task was 25% faster.

## What memory v2 changes

- The old memory, v1, preferred to keep nothing. On Sol, Codex kept a memory of only 9 to 20% of its sessions. v2 writes up more than 99% of them, and puts a summary of at most 10 KB in front of every task.
- v2 passed 66.3% of the tasks, against 49.7% for v1. That is 16.7 points more, or 100 more tasks.
- v2 reads more on every task: 184k input tokens, against 154k for v1. That is 19% more, so subscription limits run out faster.

## What past.dev adds on top of v2

- With past.dev added to memory v2, Codex passed 85.8%. That is 19.5 points more than v2 alone.
- It reads 109k input tokens per task: 41% fewer than v2 alone, and 29% fewer than v1. Each passed task costs 54% less than with v2 alone.
- v2 and past.dev work well together. For the first time, Codex's own memory adds to past.dev: with v1 added, past.dev stayed at 81%, and with v2 added, it went up to 85.8%.
- It is faster. The median task takes 30 seconds instead of 40 with v2 alone, and the slowest 10% of tasks finish under 57 seconds instead of 65.
- Codex's memory has a limit. It keeps the newest 256 conversations by default, and 4,096 at most. Because v2 keeps almost every session, the 1,000 oldest sessions of two of our three people fell out of the memory. past.dev has no limit.

## Install the Codex plugin

In a terminal:

```
codex plugin marketplace add pastdotdev/codex-plugin
codex plugin add past@past
```

Then connect the plugin to a past.dev project, as the [Codex plugin docs](/docs/plugins/codex) describe. Sign up with a work email and you get 400,000 free credits.

## How we tested

Codex got months of work history from three people, then 600 tasks that depend on it, such as reviewing a pull request or sending an email. For each memory, Codex built the memory itself from the history, on the model under test. Then Codex ran the tasks with that memory, then with the past.dev plugin, version 0.2.0, added on top. The memory did not change during the tests. The benchmark's own judge graded every task, and we ran the tests ourselves. We also ran Codex on GPT 5.6 Luna.

| | Codex on GPT-6.1 Sol | Codex on GPT 5.6 Luna |
|---|---|---|
| Tasks passed, memory v1 | 298 / 600 (49.7%) | 209 / 600 (34.8%) |
| Tasks passed, memory v2 | 398 / 600 (66.3%) | 242 / 600 (40.3%) |
| Tasks passed, memory v2 + past.dev | 515 / 600 (85.8%) | 394 / 600 (65.7%) |
| Input tokens per task, v1 / v2 / v2 + past.dev | 154k / 184k / 109k | 229k / 220k / 99k |
| Cost per passed task, v1 / v2 / v2 + past.dev | $0.138 / $0.119 / $0.055 | $0.030 / $0.026 / $0.008 |
| Median / p90 time per task, v2 / v2 + past.dev | 40 s / 65 s, 30 s / 57 s | 28 s / 139 s, 17 s / 32 s |

Dollars are at list price, because every run used a subscription.

*The tasks come from DolphinBench, Mem0's open benchmark for agent memory. Mem0 has not run or checked these results. Our earlier results on it are in the [DolphinBench results post](/blog/dolphinbench-results).*