From your conversations with AI, agent work, external sources, and more, structure and accumulate your and your team's decisions, state, and relationships — organized and kept current automatically.
The memory platform for the age of AI agents, turning them into knowledge assets the next agent can use right away.
We'll send your invite once it's ready. If requests are high, there may be a wait — thank you for your patience.
Struggling with AI memory?
If you use Claude, ChatGPT, Cursor, or Gemini every day, you keep hitting three walls.
Even as ChatGPT and Claude add memory, it stays inside each tool. Switch tools or projects and you often re-explain your structure and past decisions all over again.
Do I really have to re-explain to Cursor what I decided in Claude?
The design decisions and research you built up with AI live only in your own chat history. Teammates can't see them, and if you're away, that knowledge disappears entirely.
Only they know that spec — guess I'm researching it again.
Pasting your README, specs, and past threads every time inflates the token count and your bill — and you still can't tell how much got through.
I pasted so much context my API bill doubled.
SOLVED WITH Kagura AI
The more you use it, the more Kagura AI organizes automatically — growing into a knowledge base your whole team can use.
Capture your AI conversations, agent activity, files, and external sources via MCP, REST API, or auto-sync. You never have to decide what to keep.
It merges duplicates, links related memories, and fixes stale or conflicting information. The more you use it, the more accurate your knowledge base becomes.
Share the same memory across multiple AIs like Claude, Cursor, Codex, and Gemini. With MCP, CLI, and REST API support, continue with the same context from any tool.
One person's insight becomes the whole team's asset. Configure which memories are shared and who can see them, by role.
FOR WHOM
Whatever your AI experience level, Kagura AI helps in situations like these.
If you write code with Claude Code or Cursor: stop re-describing your project on every new chat, and stop re-teaching last week's bug fix to the AI.
Your team uses ChatGPT or Claude, but no one can see who decided what. No more digging through Slack to find how a bug was handled — the whole team uses AI well.
If you're unsure what to tell the AI or never feel like you're using it well: context accumulates automatically from past conversations, so you stop re-explaining the basics.
Want to roll out ChatGPT but worried about leaks, or teams using it in silos? See who uses what, and separate accessible information by department.
Keep client A's and B's requirements from blurring together. Separate memory per client and reuse past proposals and design decisions on the next job.
Want to pass what agent A found to agent B, or share one context across agents? Embed Kagura AI as a shared memory all your agents can reference.
FEATURE LIST
Kagura Memory Cloud saves and retrieves your AI's memory. Start free, and as you move up a plan, more of the upkeep, analysis, and team sharing happens automatically — all on the same platform.
A cloud home for your AI's memory
Just send work notes and decisions from AI tools like Claude, Cursor, or ChatGPT — they're organized and stored automatically. Connect via MCP, CLI, or REST API, ready to recall from the next session or another tool. Everyday features work on every plan; higher plans raise your quotas and add automatic upkeep, analysis, team sharing, and connectors on top of the same memory.
Hand over a conversation or note and it decides what matters, then organizes and stores it. No need to think about what or how to save.
Pull up related memories from vague phrases like 'that thing last week' or 'the login design decision'. It searches by meaning, even when keywords don't match exactly.
Automatically links the relationships between saved memories. It records context like 'this bug came from that design decision', so you can pull related info together.
Keep memory separated by project or client. No more mixing up client A with client B — hand the AI only the memory that's relevant.
Access the same memory from different AI tools like Claude, Cursor, Gemini, and Codex. Connect not only via MCP but also CLI and REST API — the same context works from any tool, any method. The REST API is available on Starter and up, and the public API on Pro and up.
Set who can see which memory. For team use, control things like 'this memory is mine only' or 'this project is open to the whole team'. Operation history is kept too.
COST BENEFITS
All figures below are simulations. Actual savings vary by project size, usage, and the AI model you use. Please treat them as pre-adoption reference only.
Sends only relevant memory instead of full context, cutting tokens by ~72% while keeping your AI precisely informed.
Assumptions & method
Accuracy drops as context grows. Delivering only relevant memory avoids the degradation confirmed across the latest generation of models.
Assumptions & method
The more you log decisions and work, the less you re-explain. Personal memory builds up into a shared team asset over time.
Assumptions & method
| Axis | Without Memory Layer (full send) | Kagura AI (Compile Layer) | Difference (simulated) |
|---|---|---|---|
| Tokens per call | ~25,000+ | ~7,000 or fewer | ~72% less |
| Monthly API (GPT-6 Sol, 1 person, 100/day) | ~$150/mo | ~$42/mo | ~$108 saved |
| Monthly API (Claude Opus 5.5, 1 person, 100/day) | ~$300/mo | ~$84/mo | ~$216 saved |
| Cost at 3× scale (3× context) | ~3× (~75,000 tokens/call · ~$450–$900/mo) | Roughly flat (~7,000 tokens/call · ~$42–$84/mo) | ~91% less (~$408–$816/mo saved) |
| New session startup | Explain from scratch | Starts with memory | No explanation cost |
All of the above are Kagura AI's own simulations and do not guarantee actual results. While modern AI context windows have expanded to 200K–1M tokens or more, academic research consistently shows that feeding long context directly into models reduces accuracy. NoLiMa (ICML 2025) found that at 32K tokens, many models fall below 50% of their short-context baseline, and long-context accuracy degradation has also been reported for later-generation large models. Monthly cost estimates use the providers’ published list input prices as of September 2026 — GPT-6 Sol: $2.00 / 1M tokens, Claude Opus 5.5: $4.00 / 1M tokens — excluding limited-time promotional pricing, batch/caching discounts, and output-token costs; other models are priced differently. Actual savings vary widely by project size, call frequency, and the nature of the context. We recommend validating in your own environment before adoption. The 3× scale estimate assumes full-context tokens per call grow in proportion to the context (~25,000 → ~75,000), while Kagura AI passes only the relevant memories and stays roughly flat at ~7,000 tokens per call (monthly figures span GPT-6 Sol to Claude Opus 5.5 list prices). Figures, rates, and sources on this page are current as of September 2026.
Primary source: Modarressi et al. (2025), 'NoLiMa: Long-Context Evaluation Beyond Literal Matching' (arXiv:2502.05167 / ICML 2025) / input prices: OpenAI API Pricing (GPT-6 Sol) and Anthropic Claude API Pricing (Claude Opus 5.5), both list prices as of September 2026
PRICING
Begin with the invite-only Trial. Move to Starter / Pro / Pro Max / Custom as your needs grow.
Available to invited users only
Request an inviteWe'll send your invite once it's ready. If requests are high, there may be a wait — thank you for your patience.
Monthly subscription · cancel anytime
Monthly subscription · cancel anytime
Monthly subscription · cancel anytime
| Trial | Starter | Pro | Pro Max | Custom | |
|---|---|---|---|---|---|
| Contexts per workspace | 5 | 30 | 200 | 1,000 | Custom |
| Memories | 500 | 3,000 | 20,000 | 100,000 | Custom |
| Memories created / day | 50 | 300 | 2,000 | 10,000 | Custom |
| MCP calls / day | 1,000 | 10,000 | 50,000 | 250,000 | Custom |
| Memory analyses / day | — | — | 3 | 15 | Custom |
| Reranking | — | ✓ | ✓ | ✓ | ✓ |
| MCP app credentials (OAuth) | ✓ | ✓ | ✓ | ✓ | ✓ |
| File storage | 100MiB | 1GiB | 10GiB | 50GiB | Custom |
| Members per workspace | 1 | 1 | 10 | 50 | Custom |
| Owned workspaces | 1 | 1 | 3 | 20 | Custom |
| Resource tokens | — | 3 | 30 | 150 | Custom |
| Connector seats | — | 3 | 10 | 50 | Custom |
| REST calls / day | — | 1,000 | 5,000 | 25,000 | Custom |
| Public API calls / day | — | — | 1,000 | 5,000 | Custom |
| Bound public API / min | — | — | 100 | 500 | Custom |
| Sleep-mode contexts | — | — | 3 | 15 | Custom |
| Shared contexts | — | — | ✓ | ✓ | ✓ |
| Public contexts | — | — | — | ✓ | ✓ |
| Resources | — | — | — | ✓ | ✓ |
| Connectors | — | — | — | ✓ | ✓ |
| Memory analysis | — | — | ✓ | ✓ | ✓ |
Included in every self-host edition
For individuals and teams with technical infrastructure know-how. No support from Kagura AI, Inc.
View on GitHubFor companies with technical infrastructure know-how. Online support from Kagura AI, Inc.
Contact usFor any company. Kagura AI, Inc. handles initial setup, maintenance, and support, with bespoke customization available.
Contact usFAQ
Stuck? Answers to your questions.
Kagura AI is a memory platform for AI agents. It's a shared place where AIs like Claude, ChatGPT, Cursor, and Gemini remember your team's work logs, decisions, and state — storing and organizing memory so any tool can recall the same context and pick up where you left off.
Yes. Even though it's a memory platform, the basics work like chatting: tell Claude or ChatGPT 'remember this' to save, and ask 'what did we decide about that?' to recall. MCP and API connections are available too, but we recommend starting with the trial plan from the screen first.
ChatGPT's memory lives only inside ChatGPT. Kagura AI is an independent memory platform that doesn't belong to any one tool, so multiple AI agents — Claude, Cursor, Gemini, Codex — share the same memory. It also adds team sharing, per-project separation, and reference history.
RAG retrieves documents from a large set and hands them to the AI. As a memory platform, Kagura AI structures not just documents but the state and history of when, who, and how something was decided. It can complement RAG.
Data on the platform is private to you by default. Only memory you set to team-shared becomes accessible to permitted members. For enterprise use, you can self-host the platform on your own servers.
Yes. We currently offer an invite-only trial plan for free — up to 1,000 memories and 1 workspace — to try the platform's core features. Just reach out via 'Request an invite'.
/compact summarizes the current conversation into a file. Kagura AI further structures that summary and stores it on the platform as memory you can search and reuse from future sessions and other AI tools.
It depends on your plan. Trial and Starter allow 1 member, Pro up to 10, and Pro Max up to 50. For larger scale or company-wide rollout, talk to us about the Custom plan.
Yes. The Slack connector (beta) links your channels and automatically turns their conversations into searchable memory, with personal data redacted before storage. For GitHub, a Source Ref feature links memory to Issues and PRs, and other tools connect via API/MCP. With the platform at the center, context scattered across tools is consolidated in one place.
Join the waitlist from the 'Request an invite' button — we'll send your invite to the free Trial plan as soon as it's ready. If you're considering team or enterprise adoption, talk to us via the 'Business inquiries' form.
Yes — you can use the self-host edition. It lets you run Kagura AI on a private server you provide, so you decide where data is stored and set your own security, compliance, and internal policies, and you can scale your infrastructure without limits. Because it's built on open-source code, the system stays durable and independent. There are three plans: OSS (free, self-operated, with no support from us), Self-host Support (with our online support), and Enterprise (we handle initial setup, maintenance, and support, with bespoke customization available). See the 'Self-host' tab under Pricing, or reach out via the 'Business inquiries' form.