NextFloor
Home Assistant용 3D 평면도, 내장 편집기로 집을 그려 네온 3D 화면에서 제어, 브라우저에서 체험
Claude Code에 지속 메모리를 제공해 프롬프트 전후 결정을 기억하고 도구 간에 공유합니다
Give Claude Code a persistent memory. Atlaso recalls your decisions before each prompt and captures them after each turn. Automatic, and shared across tools.
플러그인Python · ShellGitHub @atlaso-labs온라인운영자 미확인
써 보셨어요? 한 번만 눌러 주세요. 로그인은 필요 없어요.
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 확인하기Long-term memory for Claude Code that follows you to Codex, Cursor and your other machines: recall before every prompt, capture after every turn, a skill and Ambient Memory (Pro). By Atlaso, the AI memory layer.
저장소 README에서 가져옴
README 전체를 불러오는 중입니다.
README 전문 보기 ↗Claude Code forgets everything the moment a session ends. Atlaso remembers — it recalls your decisions, conventions, and context before every prompt and captures what's new after every turn. Fully automatic; nothing to type.
Claude Code remembers a project with CLAUDE.md and its own auto memory, and both stay on one machine and inside Claude Code. This connector gives Claude Code a memory that also shows up in Codex, Cursor and your other tools, and on your other computers on Pro: the decisions behind the code, your conventions beyond CLAUDE.md, where you left off, and how you like to work. Recall runs before every prompt and capture after every turn; you never type "remember this". Free covers one device and one active tool. How it works · Docs
/plugin marketplace add atlaso-labs/claude-code
/plugin install atlaso@atlaso
What you get
Links: Why Atlaso for Claude Code · Setup guide · What is an AI memory layer? · Dashboard
The Claude Code connector for Atlaso memory. Lives under platform/tools/<tool>/
— one folder per tool we support (Claude Code first, then Claude Desktop, Codex, …).
It's deliberately thin: every hook just calls the tool-agnostic core
(platform/client → atlaso_client.Client). Nothing smart lives here; the engine
stays on the server.
Claude Code's event names are fixed; we brand our side (script names + the block we inject). Each hook → one Atlaso action:
| Atlaso hook | Claude Code event | What it does |
|---|---|---|
recall (hooks/recall.sh) | UserPromptSubmit | Inject recalled memory before each prompt. Synchronous, fast. |
capture (hooks/capture.sh) | Stop | Save the just-finished exchange. Instant local write (no network), synced later. |
start (hooks/start.sh) | SessionStart | Inject eligible personal + current-project Ambient Memory, then sync in the background. |
end (hooks/end.sh) | SessionEnd | Flush this session's memories up + pull new ones. Detached. |
Injected block (no instructions — the model decides how to use it):
=== Atlaso Memory ===
- recalled note
- recalled note
=== Atlaso Memory ===
Hooks fail open: a missing interpreter or any error → silent no-op, the turn proceeds. A light commodity filter skips trivial acks ("ok", "thanks") on capture; the real "worth keeping" gate runs server-side.
tools/claude-code/
hooks/ recall.sh · capture.sh · start.sh · end.sh (shims)
atlaso_cc/ recall.py · capture.py · sync.py · transcript.py · _shim.py
tests/ test_recall · test_capture · test_transcript
Later: .mcp.json (active memory tools), a SKILL.md, and a plugin.json to
bundle all of it as the installable atlaso plugin.
Hooks resolve the SDK venv (it has httpx) and the atlaso_client package by
relative path. Run the tests:
cd platform/tools/claude-code
PYTHONPATH=".:../../client" ../../sdk/.venv/bin/python -m pytest tests -q
Manual end-to-end (offline, throwaway dir):
export ATLASO_GLOBAL_PATH=/tmp/atlaso-hooktest # redirects auth + cache here
printf '{"type":"user","message":{"content":"always use pnpm not npm"}}\n' > /tmp/t.jsonl
echo '{"transcript_path":"/tmp/t.jsonl"}' | ./hooks/capture.sh
echo '{"prompt":"which package manager"}' | ./hooks/recall.sh # → injects the block
ATLASO_DEBUG=1 writes per-hook logs to <atlaso dir>/atlaso-*.log.
When enabled on an eligible plan, the start hook requests a brief for this tool and the host event's project directory. Personal context can accompany the current project's context. Missing or untrustworthy project identity requests personal context only. A first session can receive a brief without a local cache. Each session validates the current server policy; disabled, offline, rejected, or incompatible responses inject no Ambient Memory. Reopening after recovery retries automatically. Per-turn recall remains a separate feature.
The Python start path has a three-second budget on POSIX; the host hook manifest sets a five-second timeout, including runtime startup. The latter is also the Windows bound. First-time runtime installation can exceed that budget and yield no brief until dependencies are available. These source changes require publishing the rebuilt plugin and updating the installed version; a source checkout alone does not update a user's plugin. Host versions must actually execute SessionStart.