# co rem — wake up with the context you need co rem is named for rapid eye movement sleep. It turns mail and local coding sessions you approve into a private notebook about people, projects and skills. When you return, it brings forward what changed, what needs attention and the original sources behind its claims. The opt-in 1.9.0a40 preview maps the sources first, investigates eligible pages concurrently, and shows the full memory and its sources in the local reader. Skill usage now says when a number is only a coding-session name match; retained evaluation attempts are counted separately, and neither count proves that the installed version ran or the task succeeded. The latest [preview notes](/releases/1.9.0a40) describe the bounded phone and desktop checks. Earlier preview changes remain in the [release archive](/releases/archive). The reader image on this page is a real product frame made with an invented notebook. No private correspondence is in the example. [Explore the full-size sample reader](/rem/demo) to follow a page, inspect a source record, and reveal an older memory before installing the alpha. The sample is frozen on 2 October 2026 and uses only invented people and sources. ## Start the opt-in preview ```bash python -m pip install --upgrade 'connectonion==1.9.0a40' co auth google # or: co auth microsoft co rem init --days 5 # a smaller first-run trial co rem open # read a local, offline snapshot co rem start # review sources and approve a schedule ``` `co rem init` maps mail and local coding sessions, then uses the configured model to investigate your page, eligible people, queued projects, related organizations and installed skills, including in scripts and JSON runs. It estimates the work before spending tokens. The configured investigation budget (35% of a weekly runner allowance by default) is a planning target. New pages can start beyond that target until the configured weekly safety floor (90% by default); pages already running finish. The input estimate includes cached tokens and is not the weekly quota meter. The first run selects projects from the map it just showed. Session folders found afterward remain candidates and do not silently add project pages to that count. `co rem start` shows which sources, runner, model, schedule and limits will be used and asks before enabling background updates. Written project pages compare the owner's project-specific coding requests with fixed, bounded local repository evidence. The current supported finding leads, partial session coverage is marked, and every numbered source can be opened. An old README documents its own snapshot, and a local commit does not prove a passing test or public release. Skill pages keep the mapped invocation name, distinguish the intended result from inspected work, and link local output files beside concrete findings when those files are available. A run report alone does not prove its output quality. Busy Gmail weeks are split without fetching the same message headers twice; requests now have a bounded timeout and retry. `co rem stop` turns the schedule off without deleting the notebook. Run `co rem open` again to render a newer snapshot. The reader runs locally and works offline. It inherits the a12 fix for Windows snapshot opening. ## What the morning reader can show - Up to three pages touched by the latest pass, with their current statement and a route to the page and its sources. Cited Facts and Contact value changes are shown separately from page rewrites. - Direct page links, unambiguous name mentions and backlinks, each labeled by its basis. A mention helps navigation; it does not prove a relationship. - A question from an older page. You can pause before revealing what that page already says; the reveal is not a learning score. - Open threads: what is waiting on you and what others owe you, with a complete action view and direct record links. - Focused people and project records with dated activity, compact facts, decisions when recorded, and the full memory and sources visible on the page. Project purpose citations stay reachable beside the purpose preview. - Skill records with openable cited run parts and local outputs when available. A cited lead finding has direct, phone-sized links to its numbered sources. Usage counts match skill names in coding sessions; separate retained eval attempts may differ, and neither verifies the installed version or outcome. - Archived original excerpts and cited conversations when the source bodies are available locally. Long source dialogs offer Find and Next match inside the visible original, including a notice when the retained excerpt is truncated. - The last pass's counts and run history, available in a disclosure. Completed nightly passes can compare cited, keyed Facts and Contact values before and after the run. Other prose changes remain page-level, and older run logs cannot be compared retroactively. Revealing an older answer still does not measure retention. [Broader claim comparison](https://github.com/openonion/connectonion/issues/2096), [durable recall](https://github.com/openonion/connectonion/issues/2097), and [source-backed questions and approved actions](https://github.com/openonion/connectonion/issues/2071) are open work. The [product audit](https://github.com/openonion/connectonion/blob/main/docs/design-evidence/rem-product-audit-2026-10-01.md) tracks the ten largest design and product problems. ## Sources and storage co rem can use connected Gmail and Outlook mail and local Codex and Claude Code sessions. WhatsApp chats are optional and selected by chat. The notebook lives on your machine under `~/.co/rem`; its readable pages are Markdown, and raw source material stays in the notebook's private state. Pages link to source records so you can inspect why co rem wrote a claim. The CLI reference describes [the exact source, review, and budget controls](/cli/rem). ## Stable channel The default `pip install connectonion` channel remains stable 1.8.10. Its experimental memory command is `co wiki`. The `co rem` alpha requires the exact version pin above. Preview releases can change before 1.9.0 is stable. Read [release channels](/releases) and the [1.9.0a40 notes](/releases/1.9.0a40) before installing.