Kimi Memory Space: What It Remembers and How to Manage It

Last verified: August 29, 2026

Kimi Memory Space is Kimi’s cross-conversation personalization feature. It can selectively save information such as your identity, stable preferences, work habits, preferred response style, and things you want Kimi to avoid, then apply relevant memories in later conversations.

It does not create a complete archive of everything you have ever said, and it is not the same as the context inside one chat. You can ask Kimi to add, update, forget, or clear memories, inspect saved entries through Settings, and turn the feature off when you do not want cross-chat personalization.

The safest way to use Kimi Memory: save short, stable preferences that remain useful across many chats; keep temporary work, evidence, passwords, private records, and project source material somewhere more appropriate.

Kimi Memory Space at a glance

The official Kimi Memory Space documentation describes a selective memory system that can carry useful personal context between conversations.

Kimi Memory Space overview showing identity, preferences, work habits, communication style, likes and dislikes, and stable constraints that can carry across conversations.

Question	Current Answer
Does Kimi remember information across chats?	Yes. Selected memories can influence later conversations.
Does it save every message?	No. Kimi says a separately trained model selects information considered useful.
What types of information may be saved?	Identity, preferences, habits, communication style, recurring work patterns, and dislikes.
Can the user add a memory directly?	Yes, through natural-language instructions such as “Remember that…”
Can memories be reviewed?	Yes, in Settings → Personalization → Memory Space or by asking what Kimi remembers.
Can memories be updated or deleted?	Yes, through conversation commands; individual entries can also be removed in Settings.
Can all memories be cleared?	Yes. Kimi documents the “Clear all memories” instruction.
Can memory be turned off?	Yes. Kimi says the feature can be disabled.
Current capacity	Up to 50 entries, with up to 500 characters per entry.
Is Memory Space the same as Kimi API memory?	No. The Direct API is stateless unless the developer adds a memory system.

Five systems commonly confused with Kimi Memory

The word “memory” is often used for several different mechanisms. Knowing which one is active prevents incorrect expectations about what Kimi will recall.

System	What It Stores or Carries	Persistence Boundary
Current chat context	Messages, files, tool results, and instructions in the active conversation	Limited by the model and conversation context; starting a new chat creates a fresh session
Chat history	Saved conversation records that the user can reopen	A stored chat is not automatically the same as a selected Memory entry
Kimi Memory Space	Selected personal facts, preferences, habits, and response instructions	Can influence later conversations until updated, deleted, cleared, or disabled
Kimi Projects	Project files, chats, and project-specific instructions	Scoped to a persistent project workspace
Kimi API or agent memory	History or facts stored by the developer, an official tool, or an agent runtime	Controlled by the application rather than Consumer Memory Space

For the amount of active information a model can process in one request, see the Kimi context-window guide. A large context window is working space, not automatic long-term memory.

What does Kimi choose to remember?

Kimi says it uses a separately trained model to decide whether conversation content is valuable enough to save. The selection process focuses broadly on two questions: who the user is, and what that user cares about, prefers, dislikes, or wants to avoid.

Useful candidates include:

Identity: your preferred name, profession, or durable role.
Expertise level: whether explanations should be introductory or technical.
Language: the language or regional variant you normally want.
Output style: concise answers, detailed examples, tables, citations, or a particular structure.
Recurring tools: a programming language, software stack, or file format you use regularly.
Stable constraints: accessibility needs or formatting rules that apply to most work.
Dislikes: emojis, excessive repetition, unnecessary disclaimers, or a tone you do not want.

Information can be saved automatically when Kimi judges it useful, but an explicit instruction is more reliable when the detail matters:

Remember that I write technical documentation for experienced developers.

Please remember that I prefer concise answers followed by a detailed appendix.

From now on, use Python examples unless I explicitly request another language.

After saving, inspect Memory Space rather than assuming the sentence was stored exactly as written.

What should not become a Kimi memory?

The official page says Kimi does not normally save unauthorized private information such as health details, passwords, or addresses unless the user explicitly asks. That safeguard should not be treated as permission to store sensitive information deliberately.

Information Type	Recommended Treatment	Reason
Password, recovery code, or API key	Never store in Memory Space	Use a password manager or secret-management system
Payment-card or banking information	Do not store	Highly sensitive and unnecessary for personalization
Home address or identity-document number	Do not store unless strictly required by a trusted workflow	Creates unnecessary privacy risk
Health or medical information	Avoid or minimize	Sensitive personal information may require stronger controls
Confidential client information	Keep in an approved project or business system	Global personalization is the wrong scope
Temporary deadline	Use the current chat, Project, or task system	It becomes stale quickly
Current project price or legal term	Keep in a versioned file	Memory is not an authoritative evidence store
Stable output preference	Good Memory candidate	Reusable across many conversations

A simple test is: Would this information still be useful and accurate in three months, across several unrelated chats? If not, it probably belongs in the active conversation, a Project, a file, or a task manager instead.

Manage Kimi Memory with the four-step cycle

1. Record a narrow, durable memory

Save one clear idea per entry. Narrow memories are easier to inspect, correct, and delete.

Weak Memory	Better Memory
“Remember everything about my work.”	“I write English API documentation for professional developers.”
“Always make everything good.”	“Use a direct answer first, then provide implementation details.”
“Remember my website project.”	“When discussing my Kimi site, use English for publishable content and Arabic for explanations.”
“I hate bad writing.”	“Avoid repetitive introductions and generic marketing language.”

2. Review what was actually saved

On the web or mobile app, open:

Settings
→ Personalization
→ Memory Space

You can also ask:

What do you currently remember about me?

List the saved memories that could affect how you answer my requests.

Which of your memories about me relate to writing style?

The settings panel is the better source for confirming saved entries. A conversational summary may combine or paraphrase multiple memories.

3. Repair outdated or incorrect memories

Use a direct correction that identifies both the old assumption and the replacement:

Update my memory: I no longer use JavaScript as my default language.
Use Python by default instead.

Correction: I am now a senior product manager, not a project coordinator.

Replace the memory about long-form answers with:
"I prefer answers under 800 words unless I request a comprehensive guide."

After updating, inspect Memory Space again. Do not assume that a new statement automatically removed every conflicting old entry.

4. Remove memories that no longer belong

Use targeted wording when only one entry should disappear:

Forget that I prefer JavaScript examples.

Delete the memory about my old job title.

Remove the memory that says I want emojis in every reply.

For a complete reset, Kimi documents:

Clear all memories.

Clearing everything is disruptive because useful personalization must be rebuilt. Review and delete individual entries first when only a small part is wrong.

How to turn Kimi Memory off

Kimi says Memory Space can be disabled through its personalization controls. Open Settings → Personalization → Memory Space and use the available Memory toggle.

The official documentation describes turning Memory off and clearing Memory as separate controls. Therefore, do not assume that disabling the feature automatically deletes existing entries. When removal is the objective, inspect the Memory Space and delete or clear the stored entries explicitly.

Disabling Memory also does not delete your ordinary conversation history. Chat history, Memory Space, and account deletion are separate controls.

Current Kimi Memory limits

Limit	Current Rule	Practical Meaning
Number of entries	Up to 50	Reserve space for stable, high-value information
Length per entry	Up to 500 characters	Store concise instructions rather than full documents
Legacy entries	Pre-April 2026 entries may exceed 500 characters	They remain stored but must be shortened before editing
Oversized update	Shows “Memory update failed”	Rewrite the information as one or more shorter entries
Internal IDs	Not visible to users	Refer to a memory by its content, not an ID
Behavior at 50 entries	Not clearly documented	Do not rely on automatic replacement or eviction

The official Kimi Memory tips page repeats these limits and provides examples of response-format and personalization instructions.

Build a practical 50-entry memory budget

Kimi does not prescribe how to divide the 50 slots. The following is an optional management framework, not an official product rule:

Memory Category	Example Allocation	Examples
Identity and durable roles	5 entries	Name, profession, expertise level, primary language
Output preferences	10 entries	Tone, length, formatting, citation style, code language
Recurring workflows	10 entries	Review process, preferred sequence, approval rules
Stable constraints	10 entries	Accessibility, compliance, brand, or technical constraints
Long-term interests	5 entries	Research fields and recurring subject areas
Reserved capacity	10 entries	Future needs and temporary testing

Do not fill all 50 entries simply because capacity exists. A smaller, well-maintained Memory Space is easier to audit and less likely to introduce contradictory instructions.

How to test whether a Kimi memory works

Use a controlled test that does not leak the expected answer in the recall question.

Create a unique test entry: “Remember that my test project codename is Silver Lattice.”
Inspect Memory Space: confirm that the entry appears in Settings.
Start a new chat: this separates persistent Memory from the previous conversation context.
Ask neutrally: “What is my current test-project codename? Do not guess if it is not saved.”
Test an update: replace Silver Lattice with Blue Harbor and inspect the saved entry.
Start another new chat: ask the same neutral question.
Delete the entry: remove the codename memory.
Verify absence: in a fresh chat ask, “Do you have a saved codename for my test project? Do not infer one.”
Clean up: confirm that no test entries remain in Memory Space.

A successful test shows that one entry was saved, retrieved, updated, and removed in that account at that time. It does not guarantee perfect recall for every memory or future product version.

Why Kimi may forget or misapply a memory

Symptom	Possible Cause	Recommended Check
Kimi does not recall a statement from an earlier chat	The statement remained in chat history but was never selected for Memory	Inspect Memory Space and save it explicitly if appropriate
Kimi uses an old preference	A stale entry still exists	Update or delete the old Memory
Kimi gives conflicting styles	Two saved entries contradict each other	Review the full list and consolidate the rules
A Memory update fails	The entry exceeds 500 characters	Shorten or split it into focused entries
Kimi recalls something in an unrelated chat	The saved memory was broad enough to appear relevant	Narrow its wording or delete it
Kimi forgets details within one long chat	The active context was compressed or exceeded	Summarize the session; this is a context issue, not necessarily Memory Space
A new chat feels unpersonalized	Memory is off, the information was not stored, or it was not retrieved as relevant	Check the toggle and saved entries
A Project answer uses an unwanted global preference	A broad Memory conflicts with the project workflow	Make the Memory narrower and place project rules in Project Instructions

Kimi Memory Space vs Kimi Projects

The official Kimi Projects documentation describes a Project as a persistent workspace containing reference files, chats, and project instructions.

Information	Memory Space	Project
Preferred response language	Good fit	Repeat only when the project differs
General writing style	Good fit	Use project instructions for exceptions
Project brief	Too detailed and too scoped	Store in instructions or a reference file
Approved prices and specifications	Do not use as the source of truth	Store in dated or versioned files
Client documents	Not appropriate	Use a permitted project workspace with suitable privacy controls
Temporary milestone	Likely to become stale	Keep in the relevant project
Permanent accessibility preference	Good fit	Can also be restated when critical

Memory personalizes how Kimi works with you. Projects preserve what a specific body of work depends on. Important factual evidence should remain in inspectable files rather than being reduced to a short personal Memory.

Memory Space vs context window and context caching

Feature	Primary Purpose	Does It Persist?
Memory Space	Select stable user information for reuse across chats	Yes, until managed or disabled
Context window	Define how much active input and output the model can process	No; it is working capacity for a request or conversation
Context caching	Reduce the cost of reprocessing an unchanged leading prefix	Provider-managed cache, not semantic personal memory
Chat history	Let the user reopen prior conversations	Stored history does not mean every fact becomes Memory

Read the Kimi context-window limits guide for active capacity and the Kimi Context Caching guide for repeated-input billing. Neither feature automatically decides which personal facts should follow you into a new chat.

Is Consumer Memory available through the Kimi API?

No automatic Consumer Memory Space is attached to ordinary Direct API requests. The official multi-turn API guide states that the API is stateless: the developer must resend conversation history or another selected context with every request.

Moonshot also provides a separate official memory tool through the Formula system. According to the official tools guide, it can support persistent storage of conversation history and user preferences when a developer integrates and invokes it.

That API tool should be governed as part of the developer’s application. Do not assume it shares entries, deletion controls, or settings with the Memory Space visible in the Kimi consumer interface.

For the base SDK integration, see How to use Kimi with the OpenAI-compatible API.

Kimi Claw uses a different memory workflow

Kimi Claw is documented as having its own long-term memory and active-context behavior. Its help center provides commands including:

/memory
/memory search <keyword>
/memory export

The Kimi Claw memory-loss guide explains that a daily reset clears the active session context and recommends explicitly writing important information to Memory. Claw also exposes memory through its Workspace configuration.

Use Kimi Claw’s own documentation and controls when managing its memory. Consumer Memory Space, Claw long-term memory, and an API application’s memory store should be treated as separate surfaces unless Kimi explicitly documents synchronization between them.

Kimi Memory privacy and data controls

The Memory Space help page states that saved Memory is under the user’s control, is not used for model training, and can be disabled or cleared.

This statement is specific to Memory Space. Kimi’s international Privacy Policy separately defines prompts, images, videos, files, and generated content as User Content and says such content may be processed to provide and improve the service, including model training and optimization depending on the jurisdiction and legal basis.

Practical implications:

Do not interpret the Memory training statement as a blanket promise covering every chat message or file.
Do not store passwords, API keys, payment information, or unnecessary sensitive personal data.
Review saved entries before using Kimi for a confidential or shared workflow.
Disable Memory when cross-chat personalization is not appropriate.
Delete or clear saved entries when removal—not merely non-use—is the goal.
Review the privacy policy applicable to the exact Kimi domain, product, account type, and region you use.

The official account-deletion guide says that once account deletion is approved, conversations, files, Memory Space data, and other content enter the deletion workflow. Some minimum information may remain for legally required periods.

For the wider distinction between product surfaces, file handling, and privacy statements, see the internal Kimi privacy and safety guide.

Run a monthly Memory Space hygiene audit

Export or record important preferences: do not make one product’s Memory the only copy of essential instructions.
Read every entry: do not rely only on Kimi’s conversational summary.
Delete temporary information: remove expired trips, old jobs, completed experiments, and obsolete tools.
Correct outdated identity details: titles, locations, responsibilities, and language preferences change.
Merge duplicates: replace several overlapping style rules with one concise entry when possible.
Resolve contradictions: remove rules that tell Kimi to be both extremely brief and maximally detailed by default.
Narrow broad memories: replace “always” instructions with conditions when the rule has exceptions.
Remove sensitive data: clear information that should never have been stored.
Test one or two high-value entries: use a fresh chat and neutral recall questions.
Leave spare capacity: do not keep all 50 slots occupied with low-value details.

A good Memory Space should look like a compact personalization profile, not an unstructured biography or project database.

Frequently asked questions

Does Kimi remember every conversation?

No. Kimi says it selectively saves conversation information considered useful, such as identity, preferences, habits, and dislikes. A chat can remain in history without every detail becoming a Memory entry.

How do I see what Kimi remembers?

Open Settings → Personalization → Memory Space on the web or mobile app. You can also ask, “What do you currently remember about me?” Use the settings panel to confirm the actual saved entries.

How do I add a Kimi memory?

Use a direct instruction such as “Remember that I prefer Python examples” or “Please remember that I want concise answers.” Then inspect Memory Space to verify how the information was saved.

How do I update an incorrect memory?

State the correction and replacement explicitly: “Update my memory: I no longer use JavaScript by default; use Python instead.” Review the saved list afterward and remove any conflicting old entry.

How do I delete one memory?

Tell Kimi to forget the specific information or delete the corresponding entry from Settings → Personalization → Memory Space.

How do I clear all Kimi memories?

Kimi documents the instruction “Clear all memories.” Because this removes useful personalization as well as unwanted entries, inspect and delete individual memories first when possible.

How many memories can Kimi store?

Memory Space currently supports up to 50 entries, with each entry limited to 500 characters. Older entries created before April 2026 may exceed 500 characters but must be shortened before they can be edited.

Why does Kimi show “Memory update failed”?

The official documentation associates this message with an entry exceeding the 500-character limit. Shorten the text or divide it into separate, narrowly scoped memories.

Does turning off Memory delete existing entries?

Kimi describes disabling Memory and clearing Memory as separate controls. Do not assume that switching the feature off deletes saved entries; inspect the Memory Space and remove them explicitly when deletion is required.

Is Kimi Memory used to train AI models?

The Memory Space help page says saved memories are not used for model training. Kimi’s broader privacy policy separately states that other User Content may be used to provide and improve services, including model training depending on the applicable region and legal basis.

Is Kimi Memory the same as the context window?

No. Memory Space stores selected personal information across chats. The context window is the temporary token capacity available to the model for the active request or conversation.

Does the Kimi API use my Consumer Memory Space?

Ordinary Direct API requests are stateless and do not automatically inherit Consumer Memory Space. Developers must resend history, build their own memory layer, or integrate an available memory tool explicitly.

Official sources and update methodology

This article was verified on August 29, 2026. Memory behavior, commands, settings, limits, and training statements were checked against the official Kimi Memory Space and Memory Tips pages. Conversation context, Projects, Kimi Claw, API state, account deletion, and privacy distinctions were checked against their respective official documentation.

Memory behavior, limits, settings labels, and data policies can change. Review the actual Memory Space panel and current official documentation before relying on the feature for sensitive or business-critical workflows.

Related internal resources: learn how to use Kimi AI, understand Kimi context-window limits, review Kimi Context Caching, configure the OpenAI-compatible Kimi API, or read the Kimi privacy and safety guide.

Kimi Memory Space: What Kimi Remembers and How to Manage It