One system.
Every agent.
In sync.

Architecture, decisions and their reasons, conventions, constraints, bug root causes, dependency limits, workflows, domain rules — each with its evidence files, dates and observers. cosmos why redis.

How we write, test, point and review, in one file. Injected first into every session. Add a rule from the UI, the CLI, or by typing remember: ….

Inventory, containers, deployment and API diagrams from what the repo declares, on one navigable page (atlas.html: pan, zoom, full screen) as well as in Markdown. Hash-checked for drift. Dreams let the model follow the Atlas prompt and add system context, data flows and a dependency index.

Every fact under the feature a product manager would name — assigned by the model, refined from paths. Who is active where, where people overlap, which lanes carry open findings.

Map a feature before coding: lanes touched, colliding decisions, flares in the way, suggested owner. Saved beside the code it will change. cosmos horizon "…" or /horizon in your agent.

Tests run, file:line cited, findings addressed, self-review done — or the turn is not over. Held at most once, never loops, never gates docs.

QA findings with a lifecycle — open, claimed, PR open, fixed, needs a human, won't fix, withdrawn, regressed. Board, Slack cards, no double posting.

The model curates every new observation — keep or drop, rewrite, category, lane — then dedupe, contradictions, staleness, evidence-based supersession. Uses your Claude Code login by default; no key to paste.

The last thing said on a branch is kept as its handoff, automatically; cosmos_handoff when stopping mid-work. The next session on that branch, on any machine, opens with it.

Git history and pull-request review comments feed the ledger with nobody installing anything: every commit becomes a journal line; commit bodies and review remarks are offered to the model as candidate rules and facts.

The ledger is a conformant OKF v0.2 bundle — Google's open spec for knowledge as markdown with frontmatter — so the format is not invented here. Provenance, trust and freshness on every note; relationships as OKF links: each lane is a concept page linking its facts, flares, horizon notes, the services whose code it touches, and the people active there. Any OKF consumer or graph viewer traverses it. Retrieval is BM25 with an evidence-path boost; no vectors.

Every dream that changed the memory measures two things on the repository's own facts and code: how often the fact about the code being edited is shown before the edit, and recall@5 for questions and file lookups (only files with five facts or fewer, where one fact can rank first). A fall of five points or more is named on the Dreams page and in cosmos doctor; numbers are only compared with runs measured the same way. On two real repositories, choosing pre-edit facts by the edit raised the first number from 39% to 81% and from 52% to 85%. cosmos eval runs both now.

cosmos watch: one local process that tails every agent's session files for this repo — all worktrees, subagents — captures what is new, shows who is working on what right now, and starts dreams. No hooks needed, so it also covers sessions opened before cosmos existed.

The moment the agent opens a file to change it, it sees the open flares on that file, its explicit rules, and the facts that match the code being changed: a file can carry a hundred facts, and the one about the function being edited is the one that matters. Each fact is shown once per session, so a later edit elsewhere in the file brings what is relevant there. Knowledge arrives at the decision, not in a report afterwards.

What the team did, not only what it learned: one line per agent turn — ask, files, commits, branch — filed by lane and person. Five commits and no new fact is still on record.

cosmos ui: overview, ledger, lanes, atlas, charter, horizon, flares, dreams, verdicts, activity, docs. Localhost only. Near-black and gold over a drifting night-storm sky behind glass panels, readable on a phone. The sidebar folds to icons; a fact or flare opens in a dialog; long lists page in the browser, so thousands of facts need no server.

/recall, /flare, /qa, /horizon and seven more, the same in Claude Code, Gemini CLI, Cursor, Copilot, Windsurf and Codex, written by cosmos connect.

cosmos's work as Claude Routines, in the cloud on a schedule or a GitHub trigger, with no laptop open: pr-review against the Charter and open flares, a daily pulse, a nightly dream, a weekly scan, merge-memory, flare-triage, digest. One line to set each up; the steps come with cosmos's upgrades. cosmos routines

A team's master QA protocol, runbook or SEO playbook becomes a command in every agent, found by itself (in Claude Code it also appears in the skill list; Codex gets it as a project skill in .agents/skills/; every agent finds it listed in AGENTS.md). cosmos playbooks add ../other/docs/qa/PROTOCOL.md shares one; cosmos playbooks add qa starts from a generic QA protocol.

cosmos export: flares, facts, rules, lanes, endpoints and playbooks, one sheet each; the ⬇ Excel button in the console does the same. Standard library only; text never runs as a formula.

lane_aliases folds near-duplicate lanes (deploy-config → deploy) everywhere; cosmos lanes --propose suggests them. cosmos review --repoint follows moved evidence through git's renames.

The same block is written to AGENTS.md, so Codex, Cursor, Gemini CLI and Copilot read the Charter and the facts too, and get the same slash commands.

API keys, tokens, passwords, private keys and credentials in URLs are redacted before anything is written. Sensitive folders can be excluded, and any word listed in ~/.config/cosmos/private_terms.txt (names, companies, anything) is redacted from every capture in every repository; the list lives in your home, so it is never committed.

Hooks finish in milliseconds and exit clean. The Gate is the one deliberate hold, and it explains itself.