AI Decision System
Replayable, offline-verifiable records for AI-assisted go/no-go decisions: experiment promotions, typed model answers and human-approved actions.
아직 감지된 업데이트가 없습니다. GitHub 릴리스가 나오면 여기에 보입니다.
개인 계정의 공개 저장소 소유자는 GitHub로 로그인해 운영자를 확인하고 소개를 직접 관리할 수 있습니다. 관리자 승인은 필요하지 않습니다.
로그인한 뒤 확인 파일을 만들어 이 저장소의 기본 브랜치에 30분 안에 커밋해 주세요. 다른 브랜치나 포크의 파일은 인정하지 않습니다. 파일을 올린 뒤 아래에서 확인하면 관리자 승인 없이 관리할 수 있습니다. 자격을 유지하려면 파일을 남겨 두세요.
로그인이 필요하면 로그인 화면으로 이동합니다. 로그인 후 이곳으로 돌아와 원하는 확인 버튼을 다시 눌러 주세요.
프로젝트 폴더에서 /nomorevibe verify를 실행하면 소유권을 확인하고 소개를 직접 관리할 수 있습니다.
1분 만에 확인하기Neuroplasticity for Agents, a living memory that focuses on relevant context.
저장소 README에서 가져옴
README 전체를 불러오는 중입니다. 아래는 저장된 발췌입니다.
Neuroplasticity for Agents, a living memory that focuses on relevant context. Help your agents remember and share their work.
Augenta saves your agent's activity and project notes so you can ask about them
later. It works with Claude Code and Codex.
Connect a project and choose where its records go. A Workspace is a place
in Augenta to keep that memory. Use your own Workspace or one shared with your
team.README 전문 보기 ↗Help your agents remember and share their work across projects.
Augenta saves your agent's activity and project notes so you can ask about them later. It works with Claude Code and Codex.
Connect a project and choose where its records go. A Workspace is a place in Augenta to keep that memory. Use your own Workspace or one shared with your team.
Nothing is captured until you connect that project.
Before you connect: saved records include full chat transcripts, which can contain secrets. Every Workspace you choose gets the same records. Read what gets captured.
You need Node.js 20 or newer and an
Augenta account. Check that your agent's terminal can run
node --version. Without Node, the plugin says so once, at the start of your
first session, and then stays quiet. Your first sign-in gives you a private
Default Workspace.
Choose your app below. The setup guide also has screenshots for desktop apps.
Send each command as its own message:
/plugin marketplace add AugentaAI/augenta-plugin
/plugin install augenta@augenta
Restart Claude Code, open your project, and run:
/augenta:connect
Run these in your terminal:
codex plugin marketplace add AugentaAI/augenta-plugin --ref main
codex plugin add augenta@augenta
Start a new session in your project. Trust the Augenta hooks when asked.
Hooks let the plugin save activity as you work. If you dismiss the prompt,
run /hooks. Changes to hooks can make Codex ask again after an update.
Then ask “Connect Augenta”, or run:
$augenta:connect
Open Customize → Plugins → Add → Add marketplace. Enter
https://github.com/AugentaAI/augenta-plugin, keep automatic sync on, and
enable Augenta.
Start a new Code or Cowork task with your project folder attached. Choose
Augenta's connect skill from the / or + menu. Cloud tasks work differently;
see Cloud sessions.
Open Plugins → Add → Add a marketplace. Use
https://github.com/AugentaAI/augenta-plugin as the source, main as the Git
ref, and leave Sparse paths empty. Install or enable Augenta.
Start a new Codex or Work task with your project. Ask “Connect Augenta”,
or run $augenta:connect.
The plugin saves records on your machine first, then sends them in the background. A short wait is normal.
To change Workspaces, run connect again. It shows your current choices and asks you to pick the full set each time. See connection settings for what happens when you add or remove a Workspace.
Ask your agent “What does Augenta remember about our sign-in setup?”
Or use the recall command in Claude Code:
/augenta:recall what did we decide about sign-in?
In Codex:
$augenta:recall what did we decide about sign-in?
Usage: /augenta:recall [answer | context] <question> (or $augenta:recall in Codex).
Your agent can also use recall when a task needs past context. Each connected
Workspace is asked, and each result shows where it came from. Augenta's model
writes the answer by default, taking up to a minute. Say context before the
question to get the saved memory for your agent to answer from without an Augenta
answer-model call. If the model is unavailable or external-model access has not
been acknowledged, the plugin tries context once and your agent explains why it
is using memory instead. An administrator can acknowledge model access to enable
answers after a consent refusal.
A new Workspace may have nothing to recall yet. If recall is not available, the plugin will say so.
Recall also runs on its own, if you turn it on. Connect asks whether to turn on automatic recall for the project; it is off unless you choose it. When it is on, each prompt you submit in the project is also a recall question:
Your agent can still use the recall command whether automatic recall is on or off.
Every selected Workspace gets the full record:
| Records | What they include |
|---|---|
| Agent activity | Messages, tool actions, and results in a common format |
| Raw transcript records | The chat records written by your coding app |
| Project memory | Saved agent notes that match this project |
| Supplied documents | Text documents you attach or mention, and PDFs supplied or referenced in supported file-tool records, including generated and temporary PDFs |
The plugin removes common secret patterns from activity and project notes. Raw transcript records are structurally sanitized but are not secret-scrubbed. This means some internal fields are removed, but passwords, keys, and private text can still be sent in the chat records. Embedded document and image bytes are replaced with hash, size and media-type references in telemetry. Eligible text and PDFs are sent separately as documents. Text documents have common secret patterns removed. PDF bytes are not secret-scrubbed, and every selected Workspace receives the whole PDF. Images remain placeholders; they are not sent as documents.
Document capture starts after this checkout connects or joins with the new disclosure. Earlier transcript history is not rescanned for documents. Existing connections keep their current activity and memory capture with attachments off until you reconnect or join again. Upgrade every installed harness before enabling attachments: an older plugin sharing the queue can skip document records. This update also makes Codex ask you to trust the changed hooks again.
Text is split into parts below 512 KiB. A PDF must fit a single envelope (roughly 380 KiB of PDF bytes); a larger PDF is skipped whole. Missing, changing or unreadable referenced PDFs are skipped. The plugin reads only the recorded file paths; it does not scan folders or fetch documents from the network.
Only connect projects you are comfortable sharing with all chosen Workspaces. The audience is the union of their members: anyone with access to any selected Workspace may see the records.
Recall sends only the question text, with no files or transcript attached. Augenta records the request but keeps no copy of the question or answer. It keeps a one-way fingerprint of the question to check for retries. Short questions can be guessed from that fingerprint, so treat them as visible to the Workspace's audience. When automatic recall is on, it uses each prompt you submit as the question, so every prompt leaves such a fingerprint in each selected Workspace.
To stop both capture and recall for a project, delete its .augenta/config.json.
This does not delete records already sent to Augenta.
To pause capture across projects, set AUGENTA_CAPTURE_ENABLED=0 in the
environment that starts your coding app. This also pauses automatic recall.
Asking with the recall command still works while a project has its config file.
To stop new attachment capture while keeping activity and memory capture, set
AUGENTA_CAPTURE_ATTACHMENTS=0 (also accepts off or false). all still
captures only text and PDFs; it does not enable images. Already queued documents
remain queued for delivery.
To turn automatic recall on or off for one project, ask your agent to change
it, or run connect again and answer that question. To turn it off for every
project, set AUGENTA_AUTO_RECALL=0 in the same environment. Capture and the
recall command keep working either way.
Connect handles setup for you. Your project stores its connection in
.augenta/config.json: sign-in profile reference, environment URLs, organization
and chosen Workspaces. Sign-in tokens stay in your private global profile. See
connection settings for file details and changing
Workspaces.
Upgrading a browser-connected project from 0.10 or earlier? Run connect once in it; the changelog says why.
A browser-connected config.json holds no sign-in token and no Connector, so
you can commit it. Everyone who checks out the repository then points at the
same Workspaces. The rest of .augenta/ stays out of Git on its own. To keep the
config private instead, add .augenta/ to your repository's .gitignore.
A cloud session runs on a machine that is discarded when the session ends. Your
Augenta sign-in there lasts only for that session, and the project's connection
lasts only if its .augenta/config.json is committed (see
Share a project's setup). Each new session then signs
in and joins it.
| Where | Does Augenta run there? |
|---|---|
| Claude Code in the cloud (claude.ai/code) | No. Cloud sessions don't install plugins that your repository or your own settings turn on. The only route is your organization's server-managed settings, set by a Team or Enterprise Owner, and Augenta hasn't been tested that way. |
| Cowork in the cloud | Partially; live Experience delivery and recall are not yet verified. Connect supports a named temporary project inside the cloud container and binds its native engine session and readable transcript. Sign-in, temporary config and unshipped records disappear at task end; each new task connects again. A folder on a separate device is a different runtime. Use a local Cowork session where available until cloud delivery is verified. |
| Codex cloud | Untested. Internet access is off by default; turn it on and allow the hosts below. |
Experimental Cowork capture commands are available for a controlled native and OTLP pilot. They require an explicit task-to-project binding and the project's affirmed Workspace set. Real Cowork delivery remains unverified; the existing connect skill does not connect a split cloud/device runtime.
On Team and Enterprise plans, an Owner controls cloud execution in Organization settings → Cowork → Run Cowork in the cloud. Local sessions remain available for existing desktop deployments; availability depends on your plan and surface. Check Claude's current execution-mode guidance and organization controls.
Wherever commands run behind a network allowlist, allow augenta.ai,
auth.augenta.ai and api.augenta.ai:
| Where | Setting |
|---|---|
| Cowork, local or cloud | Organization settings → Capabilities → Code execution → Allow network egress. It applies to sessions created afterwards. |
| Claude Code in the cloud | The environment's Network access, set to Custom. |
| Codex cloud | The environment's internet access, with all HTTP methods allowed. The GET, HEAD and OPTIONS-only setting blocks sign-in, capture and recall. |
For Cowork on Team or Enterprise, an organization owner opens Organization settings → Capabilities → Code execution, finds Domain allowlist, and adds all three hosts individually under Additional allowed domains. Allow HTTPS requests, including POST, for discovery, browser sign-in, Experience uploads, and recall. Start a new Cowork task after saving the setting. Claude's network settings guide describes the owner controls.
The screenshot shows augenta.ai; also add auth.augenta.ai and api.augenta.ai.
If a host is blocked, connect tells you which one and why, and what to ask your
administrator to allow. In Cowork, start a new task after the setting changes;
existing tasks keep their original settings. Where the environment sets HTTPS_PROXY, connect, recall and the hooks
send through that proxy and trust the sandbox's proxy certificate. This needs
Node.js 22.21 or newer; older versions ignore the proxy.
Jobs that cannot sign in through a browser can use an API key. Follow the CI and service setup. Never paste a key into an agent conversation.