Kimi Claw Guide: Setup, Memory & Test Status

Kimi Claw is Kimi’s managed interface for creating or linking an OpenClaw assistant with persistent Memory, Workspace files, Skills, terminal access and HEARTBEAT schedules. It is separate from Kimi Work and uses different deployment and permission models.

CONTROLS OBSERVED / NO USABLE INSTANCE CONFIRMED / TEST NOT RUN. Controls were visible, but eligible membership and an authorized instance were not confirmed. We stopped before creating, linking or changing anything; this is not a performance result.

Kimi AI Guide is independent and not affiliated with Moonshot AI. Product facts were checked against official Kimi pages on August 5, 2026; see our Testing Methodology.

Kimi Claw at a glance

Question	Current documented answer	Important boundary
What is Kimi Claw?	A Kimi interface for creating, deploying or linking an OpenClaw assistant	It is an agent product, not a Kimi model name
Can Kimi host it?	Kimi documents one-click cloud deployment	The Help Center says this requires Allegretto or higher; verify the live plan before purchase
Can I link my own OpenClaw?	Yes, by installing Kimi’s plugin on the device running the existing instance	Linking changes that instance and account state; back up and review the installer first
What model is used by default?	Official pages conflict: the Help Center says K2.6, while Kimi’s July 22 resource article says K3	Check the displayed model after deployment; do not rely on either page alone
What persists?	Explicit long-term Memory and Workspace files are documented	Active conversation context is separate and resets daily at 4:00 AM by default
Can it run scheduled work?	Yes; HEARTBEAT and `/cron` are documented	A schedule can act later without a fresh prompt, so use harmless tests and review delivery scope
Can it use files?	Kimi documents sending, receiving, browsing and downloading Workspace files	File availability is not proof of accuracy, confidentiality or backend deletion
Can it use Skills and a terminal?	Yes; ClawHub Skills and a web terminal are documented	Community Skills and terminal commands expand the code, network and filesystem surface
Where can it be used?	Kimi documents web, Kimi mobile apps and Telegram access, plus distinct desktop and Android deployment guides	Access surfaces and deployment locations have different permissions and retention behavior
Did our independent test run?	No	Every accuracy, persistence, file and scheduling criterion remains NOT RUN

Kimi Claw versus Kimi Agent, Kimi Work and linked OpenClaw

The products overlap, but choosing the wrong one leads to incorrect expectations about files, permissions and persistence.

Product or route	Where it runs	Primary role	Boundary to remember
Kimi Agent	Kimi’s hosted Agent workspace	Multi-step cloud tasks and a principal deliverable	Standard Agent is not a persistent personal OpenClaw instance
Kimi Claw cloud	Kimi-managed cloud instance	Persistent OpenClaw assistant with Memory, Workspace, Skills and scheduled tasks	Kimi controls the managed deployment; current plan and retention rules apply
Linked existing OpenClaw	Your existing host or device	Use an already self-hosted OpenClaw through Kimi	Kimi’s plugin is added to an environment you operate; Kimi cloud test results would not automatically apply
Kimi Claw Desktop	Your Windows or macOS computer through Kimi Desktop	One-click local OpenClaw deployment	Kimi documents a one-instance-per-account limit and local migration choices
Kimi Claw Android	An Android device	A remotely directed phone-based OpenClaw with device automation	It requires broad device permissions; Kimi recommends an idle phone and documents restricted app categories
Kimi Work	Kimi desktop Work mode	Permission-bounded local knowledge work across folders and computer tools	It is not another name for Claw and has its own permission modes and task lifecycle

Use Kimi Work when the core job is a bounded local folder or document workflow. Use standard Kimi Agent for a hosted one-off task. Consider Kimi Claw only when persistent Memory, a long-lived Workspace, Skills, external chat access or recurring automation is central to the job.

Cloud deployment, linking and local deployment

One-click cloud deployment

Kimi’s Claw overview says the cloud route starts at kimi.com/bot: sign in, choose Create, wait for automatic setup, then customize the name and persona. It says one-click deployment requires an Allegretto or higher plan and connects Kimi membership benefits and Web Search without separate API configuration.

Do not interpret one-click as zero-risk. Creation provisions a persistent instance and can consume plan resources. Before clicking Create, record:

the displayed plan requirement and any credit wording;
whether the selected route is cloud, desktop, Android or link-existing;
the shown model and context settings;
the Workspace and Memory state;
every requested permission or plugin installation;
the intended test data, output and stop conditions; and
how the instance, synthetic memory and scheduled tasks will be removed later.

The official default-model pages conflict

The Help Center overview checked on August 5 says a new cloud Kimi Claw is provisioned with Kimi K2.6 and that K3 can be selected later. Kimi’s official resource article, published July 22, says the Create flow automatically configures Kimi K3. We found no dated deployment observation that resolves the conflict for this site.

Therefore this guide does not declare one universal default. After deployment, record the exact model displayed in settings or by the documented status command. Treat any result as specific to that model and date. Changing the model is a separate configuration action and must not be performed during a read-only access check.

Link an existing OpenClaw

Kimi documents a Link Existing OpenClaw route that installs the Kimi plugin on the device hosting OpenClaw. Kimi’s resource FAQ says an existing local OpenClaw can be linked without the cloud-deployment membership requirement, but access terms can change. Confirm the live screen before proceeding.

Linking is not a harmless preview. It changes an existing environment and may expose its persona, Memory, Skills, Workspace or chat capabilities through Kimi. Back up the instance, inspect the official installer route, define rollback and obtain authorization from the system owner. Never paste installer commands copied from an unverified article or chat.

Desktop and Android are separate routes

Kimi’s current desktop guide documents one-click local deployment inside Kimi Desktop on Windows or macOS, Allegretto-or-higher eligibility, and a maximum of one Kimi Claw Desktop per account. If another local OpenClaw exists, the migration screen can offer to copy persona, Memory, chat history and Workspace files. Select only data that is authorized for the new instance.

The Android guide is a much broader device-automation product. It currently documents Android 8.0 or later, at least 5 GB of available storage, a stable connection and Allegretto-or-higher membership. It requests storage and background-running permissions and can use accessibility-style actions to operate apps. Kimi recommends a spare phone and says sensitive data should be isolated. We did not install or test this route.

Memory is not the active conversation

Kimi documents two distinct layers:

Active conversation context contains the current thread. The official Memory page says OpenClaw resets it at 4:00 AM daily by default, with a configurable schedule.
Long-term Memory stores information explicitly written for reuse across conversations. Kimi says users can ask Claw to remember a fact, inspect Workspace -> AGENTS.md -> #MEMORY, or use /memory to manage saved items.

The daily reset is not the same as deleting long-term Memory. Conversely, mentioning a fact in chat is not proof that it was saved. For anything important, inspect the visible Memory entry and then probe it from a new conversation without restating the answer.

Memory accuracy has at least four separate questions:

Did the requested fact get stored exactly?
Is the current value recalled after a fresh conversation?
Does an update replace the old value consistently?
After a visible deletion, does Claw avoid claiming the removed value?

A successful visible deletion test cannot prove backend erasure. It can only show the behavior observable in the tested interface.

What Kimi says happens after membership ends

Kimi’s current Memory page says a cloud instance is retained for seven days after membership expires, that renewal within the window restores Memory, configuration and Skills, and that after seven days the instance may be reclaimed and data permanently deleted. It also says a memory export/backup feature is in development.

These are time-sensitive vendor statements. Download authorized Workspace files and maintain independent backups before expiry. Do not assume the same rule applies to a linked self-hosted instance, Kimi Claw Desktop or Android backup.

Workspace, files, Skills, terminal and HEARTBEAT

Workspace and file transfer

Kimi says Claw can receive images and documents, send files back on current plugin versions, and expose generated files through Workspace on the web. That makes a one-output synthetic fixture suitable for testing exact file content and scope.

Before any real file use:

remove credentials and unnecessary personal information;
upload copies, not irreplaceable originals;
define the only permitted output path;
record file hashes before and after;
prohibit parent and sibling directory access; and
download and inspect the artifact outside Claw before relying on it.

Skills

Kimi describes Skills as installable modules from the ClawHub library. A Skill can be useful, but it is also an additional instruction and execution layer. Review the source, requested permissions, network behavior and update path before installation. Our Kimi Skills guide explains how to compare a Skill against a plain prompt under fixed conditions.

No Skill is required for the first Claw Memory test. Adding one would make it harder to attribute a result to base Claw behavior.

Terminal

The documented web terminal is available through Settings. Kimi lists commands including /status, /ping, /new, /compact, /reset, /skills, /cron, /memory, /logs and /debug. Kimi also warns that the terminal and plugins share a connection channel, so restarting the OpenClaw gateway disconnects the terminal.

Do not run /reset, configuration edits, installers or arbitrary shell commands as troubleshooting shortcuts. Capture the current state first and understand exactly what the command changes. A plain-language request to Claw to run a command still authorizes code execution; it is not safer merely because no shell syntax was typed by the user.

HEARTBEAT and scheduled tasks

Kimi defines HEARTBEAT as scheduled background tasks for recurring reminders, data pulls and digests, and documents /cron for schedule management. Recurring automation should be tested with a harmless fixed token delivered only inside Kimi. Do not start with news scraping, external messages, account actions or file modification.

For each schedule, record the timezone, creation time, intended trigger, actual trigger, output, retries and cleanup. One successful trigger cannot establish long-term reliability.

Preregistered independent test

Current status

CONTROLS OBSERVED / NO USABLE INSTANCE CONFIRMED / TEST NOT RUN. The protocol and synthetic fixture were prepared on August 5, 2026. Signed-in Claw controls were visible, but eligible membership and an authorized usable instance were not confirmed. Creating or linking a Claw requires a state-changing action, so we did not deploy, connect or modify anything. There is no Claw output, latency, Memory, file, scheduling or reliability result to score.

Research question

Can an eligible Kimi Claw instance use a synthetic Workspace, retain explicitly saved facts across a fresh conversation, update and remove one fact, return one requested file and execute one harmless scheduled token without touching external channels?

Synthetic fixture and ground truth

The fixture contains only fictional data. Its ground truth was written before any planned run; it is not a Kimi answer.

Fixture item	Preregistered value or rule
Workspace	`Cedar Lab`
Source files	`README.md`, `rules.md`, `project-facts.csv`
Project IDs	`P-201`, `P-202`, `P-203`
Missing status	`P-203`
Total planned budget	USD 42,000
Permitted new file	`output/claw-summary.md` only
Saved codename	`cedar-lantern-284`
Saved units	`metric`
Review-day update	`Tuesday` to `Thursday`
Temporary value to delete	`coral-ridge-771`
Saved summary format	exactly three bullets
Scheduled verification token	`CLAW-HEARTBEAT-604`

The protocol also includes five false-memory probes for a color, phone number, home city, employer and payment method that were never stored. The expected behavior is to say no saved value exists, not to invent one.

Ten planned phases

Record the plan, deployment type, displayed model, instance status and Memory baseline.
Ask for exact Workspace facts and the source filenames used.
Explicitly save the five synthetic Memory cases.
Start a fresh conversation and run the five recall probes without restating values.
Update the review day from Tuesday to Thursday and probe both new and old values.
Delete the temporary checksum and probe it from another fresh conversation.
Run the five false-memory questions.
Create only output/claw-summary.md, download it and compare it with the oracle.
Create one Kimi-only scheduled task that returns exactly CLAW-HEARTBEAT-604.
After evidence capture and separate permission, remove only the synthetic schedule and memories created by the test.

No Telegram, WeChat, Feishu, Android automation, email or third-party delivery is allowed. The test stops if Claw requests a broader scope, a real account appears, an unexpected Skill or external channel is needed, or instance cleanup cannot be defined.

Planned scoring

Area	Points	What must be exercised
Workspace exact facts	10	Exact fixture facts and cited source files
Fresh-conversation Memory recall	25	Five explicitly saved items after a new conversation
Update consistency	15	New review day retained and old value not asserted as current
Deletion behavior	10	Removed synthetic value not claimed as stored
False-memory rejection	15	Five never-stored values rejected without guessing
File correctness and scope	15	Exact required content and no file outside the one allowed path
Scheduled token	10	Exact token at the recorded time and inside Kimi only

If a phase cannot be reached because access is blocked, it is NOT TESTABLE and is not scored as zero. An access blocker is never a performance failure.

Results ledger

Criterion	Status	Evidence available
Eligible plan confirmed	NOT CONFIRMED	No executable plan entitlement was verified for this test
Instance created or linked	NOT RUN	No state-changing deployment attempted
Workspace reading accuracy	NOT RUN	No Kimi output
Fresh-conversation recall	NOT RUN	No Memory entry created
Memory update behavior	NOT RUN	No Memory entry created
Visible deletion behavior	NOT RUN	No Memory entry created
False-memory rejection	NOT RUN	No prompts submitted
File round-trip	NOT RUN	No file uploaded or generated
HEARTBEAT timing	NOT RUN	No schedule created
External channels	NOT CONNECTED	No Telegram, WeChat, Feishu or other channel action
Performance score	NONE	No criterion was exercised

What this status supports and does not support

It supports only these statements:

the official sources and product boundaries were reviewed on August 5, 2026;
a synthetic fixture, oracle and scoring protocol exist locally; and
no Kimi Claw account, instance, Memory, Workspace, Skill, task or external channel was changed during preparation.

It does not support any claim about:

Kimi Claw availability for other accounts or regions;
Memory accuracy, persistence, deletion or backend retention;
file-reading or file-generation quality;
scheduled-task timing or reliability;
the security of a Skill, plugin, linked instance or chat channel;
cloud, desktop or Android performance; or
whether K2.6 or K3 is the default in a future deployment.

Security and privacy checklist

Persistent assistants should begin with less access than the final workflow may eventually need.

Use a new synthetic-only Workspace and a disposable test persona.
Do not connect real chat channels during the first run.
Do not upload credentials, payment data, private customer content or confidential documents.
Review every Skill before installation; avoid community Skills in the baseline.
Permit only one named output and retain before/after hashes.
Use a harmless scheduled token with no web, file or external-delivery action.
Back up a linked OpenClaw before installing Kimi’s plugin.
Record the displayed model, plan, instance type, timezone and version.
Separate visible Memory deletion from any claim of backend erasure.
Remove synthetic tasks and memories only after evidence capture and explicit cleanup approval.

Troubleshooting without losing evidence

Kimi’s official troubleshooting guide lists common 400, 401, 403, 429 and 500 states and provides Restart, Repair and Restore Initial Settings controls. Use them in that order only after recording the visible error, date, instance state and non-secret logs.

Restore Initial Settings is consequential. Kimi says Workspace files and Memory are preserved, while external chat-platform connections need reconfiguration, but that vendor description is not a substitute for a backup. Do not restore or reset an instance merely to complete a test.

For long-context errors, Kimi recommends /new, reviewing loaded Skills, /compact, and /reset only as a last resort. Each action changes the state being tested. Record it as a protocol deviation and never merge results from before and after a reset as though they were one uninterrupted run.

Frequently asked questions

Is Kimi Claw the same as OpenClaw?

Kimi Claw is Kimi’s managed and integrated way to deploy or connect an OpenClaw instance. A self-hosted OpenClaw remains an environment operated on your own host; linking it to Kimi adds Kimi’s plugin and interface.

Is Kimi Claw the same as Kimi Work?

No. Kimi Work is a desktop Agent for local knowledge-work workflows and permission-bounded folders. Kimi Claw is a persistent OpenClaw product with its own Memory, Workspace, Skills, terminal and scheduling concepts.

Is Kimi Claw free?

The current Help Center says one-click cloud deployment requires Allegretto or higher. Kimi’s resource article says an existing local OpenClaw can be linked without that cloud-deployment requirement. Plans and entitlements can change, so check the live screen and our Kimi Membership guide before paying.

Does Kimi Claw use K2.6 or K3 by default?

The official pages checked on August 5 conflict. The Help Center says K2.6; Kimi’s July 22 resource article says K3. Confirm the model in the deployed instance and record it with every test.

Why did yesterday’s Kimi Claw conversation disappear?

Kimi says active conversation context resets at 4:00 AM daily by default. Information explicitly stored in long-term Memory should be managed separately. A chat mention that was never saved may disappear with the reset.

How long is cloud Memory retained after membership expires?

Kimi currently says the cloud instance is retained for seven days, can be restored by renewing within that window, and may be reclaimed with permanent data deletion after seven days. Back up authorized Workspace files before expiry and recheck the current policy.

Can Kimi Claw send and receive files?

Kimi documents both directions and says generated files can be browsed and downloaded from Workspace. Exact behavior can depend on the current plugin version and deployment type, so verify with synthetic files first.

Can I connect Telegram or another chat platform?

Kimi documents Telegram and other channel integrations. A connection can expose messages and enable actions outside the Kimi web interface. We did not connect any channel, and a future connection requires separate explicit permission and a channel-specific security review.

Did Kimi AI Guide test Kimi Claw?

No. The test was blocked before deployment because eligible access and an authorized instance were not confirmed. The protocol is ready, but all product results remain NOT RUN.

Related Kimi AI guides

Official sources

The pages below were checked on August 5, 2026. Help Center pages did not display a publication or update date in the retrieved article body; Kimi’s resource article displays July 22, 2026. Product behavior and plan requirements may change.