# supermemory

> supermemory is building the default engine for memory and continual learning for agents. Available through our API, plugins, and MCP.

This is the Markdown representation of https://supermemory.ai/, kept in step with the page itself. For the API and the rest of the site, start with the machine summaries below.

- Build with supermemory: https://console.supermemory.ai
- Copy setup prompt: https://supermemory.ai/setup-prompt.md
- Talk to sales: https://cal.com/supermemory/meet?overlayCalendar=true
- Short summary for models: https://supermemory.ai/llms.txt
- Full summary for models: https://supermemory.ai/llms-full.txt
- API specification: https://supermemory.ai/openapi.json
- Pricing as Markdown: https://supermemory.ai/pricing.md
- Setup prompt for a coding agent: https://supermemory.ai/setup-prompt.md

Announcement: [Introducing supermemory](https://x.com/supermemory/status/2097035274094272935)

## Mission

The path to useful AI is to scale [in-context continuous learning](https://supermemory.ai/#q-hassabis) that works with **any model, any harness**, and for all the varieties of use cases. As open-source model adoption grows and the world becomes even more [multi-model](https://openrouter.ai/rankings), learning and memory should be separated from the model providers and be highly interoperable (hence in-context!).

[Intelligence is no longer a constraint.](https://arcprize.org/leaderboard)

Pre-training and post-training, scaling on data and compute, will continue to push models’ intelligence, but will not enable them to learn and improve in real time.

We build all the hard infrastructure parts of a scalable memory system, with [dense, interconnected, growing learnings](https://supermemory.ai/docs/concepts/graph-memory), and an understanding of time.

Our model **learner-1** helps any model continuously improve by extracting and dreaming on the context of any user, task, or tenant, and putting it on our vector-graph database. This is then brought to the model by injecting tokens into its context in real time.

Figure 1. Raw data enters supermemory. learner-1 updates, merges, infers and forgets against the memory db, and the result is injected into your agent’s context.

Works with: [ChatGPT](https://supermemory.ai/docs/supermemory-mcp/chatgpt-web), [Claude](https://supermemory.ai/docs/supermemory-mcp/claude-desktop), [Cursor](https://supermemory.ai/docs/integrations/cursor), [Hermes](https://supermemory.ai/docs/integrations/hermes), [MCP](https://supermemory.ai/docs/supermemory-mcp/mcp).

## What we do

### Memory that keeps learning.

learner-1 extracts and dreams on the context of every user, task, and tenant, stores it in a vector-graph database, and brings it back to the model by injecting tokens into its context in real time.

### Any model, any harness.

Learning and memory live outside the model providers, so they carry across models and harnesses and stay interoperable, in context, and yours.

### The hard infrastructure, done.

Dense, interconnected, growing learnings with an understanding of time, already serving 100k+ organizations and over a trillion tokens a month.

[Read more about the specifics](https://supermemory.ai/product)

## In production

Our promise is the best memory for agents: the kind people feel, not just the kind that tops a benchmark.

### Recall latency (per query, at the API)

| Operation | Server (ms) | End to end (ms) |
| --- | --- | --- |
| Search | 187 | 356 |
| Profile | 166 | 248 |

### Tokens processed: 1T+

per month

### Organizations: 100k+

tens of millions of end users

### Independent benchmarks (SWE-ContextBench, February 2026)

> Supermemory performs best overall, achieving FAIL_TO_PASS test rate of 55.95% and the highest resolution rate of 30.30%.

[Read the paper](https://arxiv.org/abs/2602.08316)

### Public benchmarks: #1

State of the art on LongMemEval, LoCoMo and ConvoMem

[See the research](https://supermemory.ai/research)

### Our benchmarks (Against markdown-based memory, on a 9,988-file corpus)

#### Cost per question: $7.79

64% less than $21.69

#### Questions answered: 81%

up from 69%

[The research runs](https://smfs.ai/research/runs)

### What a customer says

> Reduced avg response time from 40s → 12s. Using about 40–50% fewer tokens.

[Founder, Chatarmin](https://x.com/saasjesus/status/2024410743135694991)

## supermemory runs everywhere, even in air-gapped environments.

We build all the critical infrastructure ourselves. Which means we can deploy it wherever your data lives.

[Talk to us for your deployment requirements](https://cal.com/supermemory/meet?overlayCalendar=true)

1. In your data center
2. In your VPC
3. On your laptop

Certifications: AICPA SOC 2 certified, GDPR aligned, HIPAA compliant.

### Enterprise security.

supermemory is SOC 2 certified, HIPAA compliant and GDPR aligned. We will sign your contracts and DPA, and deploy air-gapped where the data cannot leave at all.

[Trust centre](https://trust.supermemory.com)

[Self-hosting in the docs](https://supermemory.ai/docs/self-hosting/overview)

## Running memory in production?

If your agents already depend on it, they should be running on the engine the benchmarks and the numbers above describe. Any model, any harness, through our API, plugins and MCP.

[Build with supermemory](https://console.supermemory.ai)

## Writing

- [An update to supermemory](https://supermemory.ai/blog/an-update-to-supermemory)
- [SMFS: making agentic retrieval 55% cheaper AND more accurate](https://supermemory.ai/blog/smfs-making-agentic-retrieval-55-cheaper-and-more-accurate)
- [Introducing Dynamic Dreaming: supermemory now connects the dots, for you.](https://supermemory.ai/blog/introducing-dynamic-dreaming-supermemory-now-connects-the-dots-for-you)

[Read the blog](https://supermemory.ai/blog)

## Careers

We’re a small team based in San Francisco, building the memory layer for every model and every harness. If that is the problem you want to spend the next few years on, we would like to hear from you.

[See open roles](https://binary.so/UeNYHHE)

## References

### 01. Demis Hassabis — CEO of Google DeepMind

> One of the things missing from today’s systems is the ability to online learn and continually learn. So, you know, we train these systems, we balance them, we post-train them, and then they don’t continue to learn out in the world like we would.

Source: Google DeepMind: The Podcast, “The Future of Intelligence” · December 16, 2025

### 02. Jensen Huang — CEO of NVIDIA

> AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereign AI. The world needs both frontier closed models and frontier open models.

Source: X, @JensenHuang · July 24, 2026

Link: https://x.com/JensenHuang/status/2080643682408321103