What is the Context Cloud?

The Context Cloud marks the transition of Supermemory from just a memory system to a full context infrastructure. It offers composable 'lego blocks' including Memory, SuperRAG, Search and Traversal, and Operations — the building blocks to build agents the way they should be, with the best, most reliable, and delightful infrastructure components.

What can Supermemory do today?

Today, Supermemory can: Parse and extract any modality of content (PDFs, images, etc.); Learn in real time about entities and create a growing graph about them (SOTA Memory); Large scale RAG (with a bunch of features in it); Creating profiles of your users for short-term understanding (User Profiles); Figuring out what your users are doing / understanding patterns and use cases (Qualitative Observability); Keep external services like Google Drive, Notion, S3, Crawled websites, etc. in sync in real-time (Sync); and more.

How does the new pricing model work?

In the new pricing: (1) Rich tokens / normal tokens — multi-modal content will be billed 2x normal token count; (2) Separate pricing for retrieval (which also makes it the cheapest managed retrieval service!); (3) Charging for operations — for all the little configuration that you can do with Supermemory; (4) Supermemory tokens — unlike LLM tokens, we only count and bill on the unique tokens we've seen, which means if it's an ongoing conversation with an LLM, we will literally only bill on the unique tokens in the entire conversation. Often times this is 5–10x less than text tokens themselves.

What is the difference between 'rich tokens' and 'normal tokens' in Supermemory's Context Cloud pricing?

Rich tokens are charged at 2x the normal token rate and apply to multi-modal content like PDFs, images, and audio. Normal tokens cover standard text. The key saving comes from how Supermemory counts tokens at all: only unique tokens across an entire conversation are billed, not every repeated token, which typically means you pay for 5–10x fewer tokens than a raw text-token count would suggest.

Should I self-host Supermemory or use the managed cloud service?

Self-hosting via Docker is available across all tiers with full feature parity to the cloud, so the decision is purely about your compliance and infrastructure preferences, not capability trade-offs. Teams with strict data-residency or on-prem requirements should self-host; teams that want to ship fast without managing infrastructure should default to the managed cloud.

How does Supermemory's memory graph differ from what a vector database gives you?

A vector database stores embeddings and retrieves by similarity score — it has no concept of relationships, contradictions, or time. Supermemory's memory graph tracks ontology-aware edges between memories, handles knowledge updates and contradictions, supports temporal reasoning, and can traverse multi-hop relationships — capabilities that flat vector storage cannot replicate by design.

Can I bring my own vector store instead of using Supermemory's default storage?

Yes. Supermemory supports pluggable vector store backends, so you can wire in Pinecone, Weaviate, Qdrant, or a custom store instead of the default. This means you keep your existing vector infrastructure investment and satisfy any compliance or performance requirements without duplicating data.

What is 'dynamic dreaming' and when should I use instant processing instead?

Dynamic dreaming (the default) groups related documents together during ingestion to form memory from coherent semantic units, optimizing for cost and quality. Instant processing (dreaming: 'instant') handles each document in isolation for immediate recall at the cost of one additional operation per document. Use instant processing for real-time chat and personalization; use dynamic dreaming for batch ingestion pipelines and periodic syncs.

How does Supermemory handle audio and multi-modal content ingestion?

Audio files — voice recordings, podcasts, and meeting recordings — are automatically transcribed via Gemini 2.5 Flash, then chunked and indexed into the memory graph with no third-party transcription service required. PDFs, images, web pages, PowerPoint files, and video are also supported natively through the Extractors primitive, all included free on every plan.

What frameworks and agent SDKs does Supermemory work with?

Supermemory has official SDKs for TypeScript and Python, plus framework integrations for Vercel AI SDK, LangChain, LangGraph, CrewAI, OpenAI SDK, Mastra, Pipecat, and Microsoft Agent Framework. It also ships an MCP server, a CLI, and plugins for Claude Code, Cursor, OpenCode, OpenClaw, and Windsurf — so it slots into whichever toolchain you are already using.

How does Supermemory's retrieval latency compare to Mem0?

Supermemory delivers sub-300ms recall. Mem0 retrieval runs at 7–8 seconds. That gap matters in production: multi-second retrieval stalls the agent's response loop in ways users feel immediately, while sub-300ms retrieval keeps the interaction feeling responsive.

What compliance certifications does Supermemory hold, and can it run inside a VPC?

Supermemory holds SOC 2 Type 2, HIPAA, GDPR, and CASA certifications, with encryption in transit and at rest plus fine-grained access controls. VPC deployment is available at the Enterprise tier, and full self-hosting via Docker is available across all tiers for teams that need data to stay within their own infrastructure boundary.

What's the fastest way to get Supermemory running in an existing agent project?

Run 'npx supermemory' for a five-minute setup, or use 'npx supermemory install-plugin' to automatically detect your environment and install all supported plugins in a single step. For framework integration, install via 'npm i supermemory' and connect through whichever SDK your agent already uses — Vercel AI SDK, LangChain, and Mastra all have native integrations.

The Context Cloud: All Your AI Infra in One API

When I first started building supermemory, I had one goal: To build the best memory system for AI. I would talk to customers, and find out that memory was not the only thing they needed - They were all setting up 7-8 different vendors at the same time.

Memory, Retrieval, Profiles, reranking, embedding, etc.

We had all of those features before already, but our positioning made it unclear what we offer, and led to people either being confused or paying too much for what they want to do.

The product kept changing and evolving, and we started doing more and more things with the product.

Today, Supermemory can:

Parse and extract any modality of content (PDFs, images, etc.)
Learn in real time about entities and create a growing graph about them (SOTA Memory)
Large scale RAG (with a bunch of features in it)
(And RAG+Memory available as a filesystem smfs.ai)
Creating profiles of your users for short-term understanding (User Profiles)
Figuring out what your users are doing / understanding patterns and use cases (Qualitative Observability)
Keep external services like Google Drive, Notion, S3, Crawled websites, etc. in sync in real-time (Sync)
... and more.

All of these have been natural progressions of how we think agents will be built - with lego blocks, or composition of multiple such layers, using them in the best way possible to get things done.

Each use case of memory is slightly different, and has different tradeoffs. A healthcare assistant needs to remember things about health, but also needs best quality retrieval, with a latency tradeoff being acceptable. Similarly, a research/coding assistant using subagents might be OK to load things into context in the explore phase.

To serve all of these use cases, we started opening up our API and stopped being opinionated. So now, supermemory is truly yours. It's your legos to build whatever you can wish.

Me and the team sat together to find what can _not_ be built with supermemory, and honestly, we could not come up with any good answer.

Why we did not change the pricing before

The only thing stopping us from going all into this non-opinionated legos-for-context-infrastructure was the fact that our pricing only had two metrics attached to it: `tokens ingested` and `search queries`

Well, in reality it was just one metric (tokens) as we knew no one would hit our extremely high limit of search queries, ever.

So, we started deciding the new pricing: It had to be composable yet easy - Unit-based but not credits, as we don't want people to worry about credits and wonder how our pricing works (We wanted to give our users full transparency)

In our new pricing, we figured it out - All usage draws from the same credit pool, and is very clear on what is offered.

Rich tokens / normal tokens - Multi-modal content will be billed 2x normal token count
Separate pricing for retrieval (Which also makes it the cheapest managed retrieval service!)
Charging for operations - For all the little configuration that you can do with supermemory, they are now counted as operations.
Supermemory tokens - Unlike LLM tokens, we only count and bill on the unique tokens we've seen. Which means if it's an ongoing converastion with an LLM, we will literally only bill on the unique tokens in the entire conversation. Often times this is 5-10x less than text tokens themselves.

Supermemory pricing cards: Memory $0.005, SuperRAG $0.001, Search and Traversal $0.005, Operations $0.10 per 1K

Depending on how you count it, this makes Supermemory the cheapest memory layer on the market. yes, headline prices didn't move. The math underneath did. Most of you will pay less for the same workload starting next month, and some of you will pay significantly less.

And now, no matter what, our customers have the perfect breakdown of how much they spend.

Daily spend bar chart with a tooltip showing 184.2K rich memory tokens, 33.6K text tokens, 86 operations and 479 queries

This, along with the launch of our filesystem (smfs.ai) marks the transition of supermemory moving from just a memory system to becoming the Context Cloud.

The legos to build your agents, the way they should be. And we provide the best, most reliable, and delightful infrastructure components that doesn't break your bank while helps you build and ship fast.

We're ready for the next chapter. are you?

Dear reader, we just made supermemory insanely cheap... the Context Cloud

Why we did not change the pricing before

Other posts.