A service that provides a “memory layer” for AI agents. Chat history, documents, PDFs, web pages, images, and audio can all be stored in a single foundation and retrieved when needed. The official site positions it as “context infrastructure for AI agents,” and its distinguishing trait is that a custom vector graph plus hybrid search turns it into more than a storage bin — it also extracts a user’s intent and preferences. Connectors for Notion, Google Drive, and Gmail, TypeScript / Python SDKs, a REST API, and an MCP server are all available, so developers can embed memory into their own apps and assistants.
Anyone who uses an AI assistant such as ChatGPT or Claude eventually hits the moment of re-explaining something already discussed. Each service has its own mechanism for carrying context across conversations, but building the same thing for your own app or your team’s documents means constructing a vector database and implementing retrieval from scratch. Supermemory takes over that entire layer.
Key Features
- Multi-format ingestion (Extractors): PDFs, web pages, images, and audio can be fed in as they are. There is no need to write your own conversion or chunking pipeline
- Structured recall through a memory graph: A custom vector graph engine links memories with ontology-aware edges, letting you follow connections between related pieces of information that plain similarity search would miss
- Hybrid search: Vector search and keyword search are combined. The official site advertises sub-300ms response times, so calling it mid-conversation does not introduce a noticeable wait
- Automatic user profiles: Accumulated behavior is distilled into a profile of the user’s intent, preferences, and context. A tendency such as “this person prefers concise answers” can live in the memory layer without being configured explicitly
- Connectors for external services: Notion, Google Drive, OneDrive, Gmail, S3, and a web crawler can be connected so existing documents become knowledge sources directly (available connectors differ by plan)
- SDKs, REST API, and MCP server: TypeScript / Python SDKs and an OpenAPI-specified REST API are joined by an MCP server, letting Claude, Claude Code, and other coding agents read and write memory directly
- Self-hosting: Running on your own infrastructure is supported, and higher plans add air-gapped self-hosting as an option
Pricing
| Plan | Monthly price | What’s included |
|---|---|---|
| Free | $0 | About $5/month of included usage, Supermemory MCP, community support |
| Pro | $19 | About $20/month of included usage, 2 seats, Google Drive / Notion / OneDrive connectors, Claude Code plugin, email support |
| Max | $100 | About $130/month of included usage, everything in Pro plus Gmail / Granola connectors, priority support |
| Scale | $399 | About $600/month of included usage, up to 10 seats, all connectors (Gmail / GitHub / S3 / web crawler), SOC 2, HIPAA BAA, self-hosting option |
| Enterprise | Contact sales | Unlimited usage, air-gapped self-hosting, SOC 2 / HIPAA / GDPR, dedicated account manager, uptime SLA |
Pro and above support automatic top-up for usage beyond the included amount, and Scale and above add spend caps. No annual plan is listed on the official pricing page.
Pricing is current as of August 2026. Check the official site for the latest details.
Pros & Cons
✅ Pros
- No need to build a memory foundation from scratch — choosing a vector database, writing chunking logic, and tuning retrieval are all handled for you
- PDFs, images, and audio all enter through the same door, so separate pipelines per format are unnecessary
- The MCP server means you can try it immediately from Claude or Claude Code, before writing any code
- The free plan comes with a usage allowance, so small-scale evaluation costs nothing
- Self-hosting is available, which keeps it viable in environments where data cannot leave the premises
⚠️ Cons
- It is fundamentally a developer-oriented service that assumes you will work with SDKs and APIs; it is not a no-code tool
- Pricing is usage-based, so the monthly bill is hard to predict depending on data volume and query counts
- Connectors are gated by plan — wanting only the Gmail integration still pushes you to a higher tier
- Compliance requirements such as SOC 2 and HIPAA start at Scale, which is a steep price for a small team
- If you want the agent itself to manage and edit its own memory, other options may fit that design better
Comparison with Similar Services
| Criterion | Supermemory | Mem0 | Zep | Letta |
|---|---|---|---|---|
| Positioning | Memory infrastructure for agents | General-purpose memory layer | Knowledge graph strong on time | Agent framework with self-managed memory |
| Strength | Multi-format ingestion and hybrid search | Extracting memories from conversations, broad framework support | Temporal reasoning about when a fact was true | The agent editing its own memory |
| Data formats | PDF, web, image, audio, and more | Mainly conversation and text | Mainly conversation and events | Mainly conversation |
| External integrations | Notion / Google Drive / Gmail / S3 and more | Various framework integrations | SDK-centric | SDK-centric |
| MCP support | Yes | Yes | ─ | ─ |
All of them share the goal of giving AI a memory, but their emphasis differs. Roughly speaking: Supermemory if you want existing documents and files turned into a knowledge source, Mem0 if you want to quickly wire in memory extraction from conversation logs, and Zep if you need to reason about how long a fact stayed valid.
Who Is It For
- Developers who want their own AI app or chatbot to carry context across conversations
- Teams who want internal Notion and Google Drive documents made searchable from AI
- Anyone who needs a foundation that searches across non-text material such as PDFs and audio
- People who want to handle memory from Claude or Claude Code through MCP
- Small teams looking to cut the effort and cost of operating a vector database themselves
Summary
Supermemory is infrastructure for giving AI agents long-term recall. Multi-format ingestion, structuring through a vector graph, and hybrid search come together to remove a large amount of the work involved in building a memory system yourself. The MCP server makes it easy to sample from Claude, though it truly earns its keep once you embed it into an app via the SDK or API. A sensible path is to load a small amount of your own data on the free plan, judge how the search feels, and then consider a paid tier.