ConnectOnion 1.8.9 · as co wiki, Experimental

co rem — your agent's memory

Name: ConnectOnion
Author: ConnectOnion

In REM sleep the brain replays the day and keeps what matters. co rem does the same for your work. It reads your mail and your Codex and Claude Code sessions, and keeps a Markdown page on each person, project and tool you work with. Your agent reads those pages, so it picks up where you left off instead of asking you again.

Today it is co wiki. It ships in ConnectOnion 1.8.9, labelled Experimental, so its commands may still change. In 1.9.0 it becomes co rem, the release that supports it long-term. Every command on this page is one you can run now.

Start

pip install connectonion
co auth google            # or: co auth microsoft — the mail it reads
co wiki init              # map 90 days of mail and sessions into pages; no model, no cost
co wiki open              # read the notebook in your browser
co wiki start             # approve what it reads, then turn on the daily update

init builds the frame: a page for each person you write to, each organisation, each coding project and each installed skill, plus your own page. Those pages start mostly empty. To fill your own page first, run co wiki investigate me --quick; it reads a bounded sample and says how much of the material it covered.

What it is for

It is written for the times you come back, not for daily check-ins:

Pick up where you left off. A project page says why you are doing it, where it stands, and what is still open.
Hand context to someone. Pages are plain Markdown, so a project page can go to a colleague or to their agent without you in the room.
Look back on a period. What you actually made, and what did not work.
Find a fact fast. co wiki search "term sheet" searches every page without calling a model.

The top of a page is short; the detail, and the source of each claim, is further down. These are the design goals (#1580). The feature is experimental, and how well each one works today varies.

What it reads

Source	How it gets in
Gmail	`co auth google`. 90 days of headers, then a private copy of each message body
Outlook	`co auth microsoft`, the same way
Codex sessions	Your local session files, read in place
Claude Code sessions	Your local session files, read in place
WhatsApp chats	Only the ones you add: `co wiki sources add whatsapp --chat <id>`

co wiki sources shows what is on. co wiki start shows the sources, the model and the schedule, and asks before anything is read on a schedule. It asks again whenever any of them change.

Where it keeps things

A folder on your machine, ~/.co/wiki: people/, projects/, orgs/, skills/ and a few more, all Markdown.
Private materials stay there. Saved mail bodies live in the notebook's .state/ folder, which only you can read (0700 folders, 0600 files). A mail body never goes into a page.
The model reads what it needs through your own login. With the default runner, your mail is read by Codex through your own Codex account, and nothing is billed to OpenOnion.
Unattended runs are confined. Codex runs with --sandbox workspace-write and Claude Code with --permission-mode acceptEdits. Either can write only inside the notebook's task folder, with no shell commands and no network.

What runs the model

By default, your own Codex plan. It spends no OpenOnion credits. The notebook reads Codex's weekly meter before and after every run:

Investigation has a budget of 10 points of your Codex week by default.
It starts no new page once your week is 70% used, so your own coding comes first.
co wiki status shows both numbers.

You can switch to Claude Code (co wiki config set runner claude-code model default) or to a model through co ai, including a local Ollama model (co wiki config set runner coai model ollama/qwen3).

What it costs, measured

Step	What we measured
`co wiki init`	About 10 minutes to map 90 days of two mailboxes. No model, no tokens
`co wiki investigate me --quick`	8–9 minutes in our runs (500 s and 536 s), covering 24 of 60 gathered items
A very large subject	Still millions of tokens: one took 8.2 million over three and a half hours

The last row is a known limit. Investigation is moving from summarising everything to searching the saved material (#1850).

Keeping it current

co wiki start installs a schedule on macOS. By default it runs at 03:00, 04:00, 06:00, 17:00, 18:00 and 19:00 local time, and a day's model calls are capped. On Linux and Windows the schedule is not installed yet; run co wiki sync yourself. co wiki stop turns it off and keeps your pages.