Supermemory: Long-Term AI Memory as a Service

6 min read

What Supermemory is: a hosted memory API with fact extraction, user profiles, hybrid RAG search and connectors. How it works, Python SDK, local mode, pricing.

Supermemory is long-term memory sold as a service: one API that extracts facts from conversations, keeps a profile for each user, and searches your documents and those memories together. You send content scoped to a user or project, and Supermemory handles chunking, embeddings, extraction, contradictions and forgetting. It also runs locally as a single binary.

This page covers what Supermemory does, how its memory model works, the current Python SDK, the local server, pricing and its benchmark claims. Sources are the Supermemory GitHub README, the docs and the pricing page, checked on 8 October 2026.

What is Supermemory?

Supermemory is a managed memory and context layer for AI applications. It ingests conversations, files and synced data, extracts facts about users, maintains a static and dynamic profile per user, and returns relevant memories and document chunks in one hybrid search call. It targets both app builders (through its API) and individual users (through plugins for AI tools).

Supermemory was started by Dhravya Shah and raised a seed round in October 2025. The GitHub repository, created in February 2024, had about 31,000 stars in October 2026.

What “memory as a service” means here

A memory-as-a-service API takes over the pieces teams otherwise build themselves: a vector database, an embedding pipeline, a chunking strategy, an extraction prompt, and rules for when facts change. Supermemory’s README pitches exactly that: “No vector DB config. No embedding pipelines. No chunking strategies.”

The tradeoff is the usual one for managed infrastructure. You get speed and less to operate. You give up control over how extraction and ranking work, and your users’ data lives with a vendor unless you take the local or Enterprise route. For the broader design space, see persistent memory for AI.

How Supermemory works

The README describes five parts behind one API:

ComponentWhat it does
Memory engineExtracts facts, tracks updates, resolves contradictions, forgets expired information
User profilesKeeps static facts (“Senior engineer at Acme”) and dynamic context (“Working on auth migration”)
Hybrid searchReturns document chunks (RAG) and personal memories in one query
ConnectorsSyncs Google Drive, Gmail, Notion, OneDrive, GitHub and a web crawler, with webhooks
File processingPDFs, images (OCR), video (transcription), code (AST-aware chunking)

Two behaviors set it apart from a plain vector store. First, knowledge updates: the README’s example is that “I just moved to SF” supersedes “I live in NYC.” Second, automatic forgetting: a temporary fact like “I have an exam tomorrow” expires after the date passes.

Memory is scoped by a namespace in the v5 API (called a container tag in older versions). A namespace can be a user ID, a project, a repo or a client.

Using the Supermemory API from Python

The Python SDK moved to a namespace-first v5 API with release 5.0.0 on 6 October 2026, so older examples using container_tag= and q= are out of date. This matches the current Python SDK README:

 1from supermemory import Supermemory
 2
 3client = Supermemory()  # reads SUPERMEMORY_API_KEY
 4
 5# Store content for one user; extraction happens on Supermemory's side
 6client.add("user_dana", content="Dana prefers short answers and works in Go.")
 7
 8# Profile: maintained facts, no query needed
 9profile = client.profile("user_dana").profile
10print([fact.memory for fact in profile.static])   # long-term facts
11print([fact.memory for fact in profile.dynamic])  # recent context
12
13# Hybrid search over memories and document chunks
14response = client.search("user_dana", query="how should I answer Dana?", limit=5)
15for result in response.results:
16    print(result.memory or result.chunk, result.similarity)

search takes a search_mode of "hybrid" (default), "memories" or "chunks", plus options for reranking and query rewriting (the reference notes rewriting adds about 400 ms). Deleting is done with client.memories.forget and client.memories.forget_matching, and documents have their own update and delete calls.

Framework wrappers exist for the Vercel AI SDK, LangChain, LangGraph, the OpenAI Agents SDK, Mastra, Agno and n8n.

Supermemory for AI tools: plugins and MCP

For people who just want their assistant to remember them, Supermemory ships open-source plugins for Claude Code, Cursor, Codex, OpenCode, OpenClaw and others, and acts as a memory provider for Hermes Agent. There’s also a hosted MCP server at https://mcp.supermemory.ai/mcp with three tools: memory (save or forget), recall (search plus profile summary) and context (inject the profile at the start of a chat). See memory for AI coding agents for how these compare with other approaches.

Running Supermemory locally

Supermemory local is a single binary: curl -fsSL https://supermemory.ai/install | bash, then supermemory-server. It sets up an embedded graph engine and local embeddings (Xenova/bge-base-en-v1.5 by default), prints an API key, and serves the same API on http://localhost:6767. You bring any LLM, including Ollama for offline text memory. Data lives in ./.supermemory.

Be clear on what this is. The local vs Enterprise docs say the binary is “free within its lite license limit,” built for individual developers, and that “its server source is not in the public repository.” It runs as one process with one API key and no connectors. URL ingestion still calls a hosted reader service. Enterprise adds proprietary extraction models, team access, observability and dedicated deployments.

So Supermemory can be self-hosted, but it isn’t open-source infrastructure in the way Mem0, Graphiti, Cognee or Hindsight are. Hindsight’s MIT release, for comparison, includes its full server: graph, temporal recall, consolidation and reasoning over memory.

Supermemory pricing

PlanPriceIncluded usage creditsNotable
Free$0$51 seat
Pro$19/month$203 seats, auto top-ups
Max$100/month$130Gmail connector
Scale$399/month$600Unlimited seats, self-hosted option, S3 and web crawler connectors
EnterpriseCustomCommitted spendAir-gapped or dedicated instance, SSO

Usage draws on the credits at one rate card for all plans: $0.005 per 1,000 SM tokens for plain-text memory, $0.010 for rich content, $0.001 per 1,000 SM tokens for text RAG, and $0.005 per 1,000 search queries. Subscription credits reset monthly; top-ups don’t expire.

Supermemory benchmark claims

The README says Supermemory is “#1” on LongMemEval, LoCoMo and ConvoMem, and reports 95% Recall@15 on LongMemEval while adding about 720 tokens of context. Two cautions:

  • These are self-reported. No independent run is cited.
  • Recall@15 measures whether the right evidence appears in the top 15 retrieved items. It’s a retrieval metric, not the end-to-end answer accuracy most other vendors report, so the numbers aren’t comparable.

On answer accuracy, Mem0’s comparison pages list Supermemory at 85.2 on LongMemEval, against 94.4 for Mem0 and 94.6 for Hindsight. That 85.2 matches one configuration in Supermemory’s own LongMemEval report (Gemini 3 Pro answering, GPT-4o judging). These are vendor-run numbers too.

Supermemory also maintains MemoryBench, an open-source harness for running Supermemory, Mem0, Zep and others on the same benchmarks. Running it on your own data is more useful than any headline number. Our LLM memory comparison lists each vendor’s claims.

When Supermemory fits

Supermemory fits when you want memory, document RAG and connectors from one API and don’t want to run infrastructure: consumer assistants, support bots that need both a knowledge base and user history, or a personal memory across AI tools. It fits less well when open, self-built infrastructure or full control over extraction is a requirement. In that case the closest fit is often Hindsight: MIT-licensed with every feature in the open-source server, and ahead on long histories (73.9 on BEAM 1M and 64.1 on BEAM 10M, vendor-reported). What you give up is Supermemory’s strongest part: Hindsight has no Drive, Gmail or Notion connectors and no document RAG, and each write calls an LLM. Those options are compared in Supermemory alternatives and, head to head, in Mem0 vs Supermemory, and the concepts behind them in AI agent memory explained.