Build AI assistants that retain context across conversations
Supermemory eliminates the need to re-explain settings in every AI interaction. It is a system where AI automatically learns user preferences and past conversations, retrieving the appropriate context when needed. It extracts key facts from conversations to build user profiles and discards information that is no longer valid over time.
A structure that unifies memory and retrieval
Supermemory processes simple document retrieval (RAG) and user memory within a single query. While previous methods only found document chunks and provided identical results to all users, this system tracks the changing state of users. For example, if the information 'I live in New York' is updated with new information 'I moved to San Francisco,' the system automatically reflects the latest state. Temporary facts are automatically deleted after their expiration date, preventing noise from accumulating. This process occurs in the background, so developers do not need to configure separate pipelines.
How to connect AI tools and apps
Developers install the client using npm install supermemory or pip install supermemory and store or retrieve data with simple methods like client.add() and client.profile(). Wrappers are also provided for direct integration with major AI frameworks such as Vercel AI SDK, LangChain, and OpenAI Agents SDK. In this case, the API runs at http://localhost:6767, and you can connect Ollama to run it completely offline. Since both cloud and local environments use the same API, almost no code changes are required when moving from prototyping to a deployment environment.
Performance metrics and benchmark results
Supermemory recorded first place in three major AI memory benchmarks: LongMemEval, LoCoMo, and ConvoMem. In LongMemEval, it achieved 95% Recall@15 while adding only about 720 context tokens, demonstrating a 99.4% context reduction effect. Accuracy by category is 99% for knowledge updates, 100% for assistant memory, 97% for user memory, 93% for multi-session, 91% for temporal reasoning, and 90% for preferences. Additionally, the self-developed Supermemory Filesystem (SMFS) was confirmed in the xAFS benchmark to use 3.0 times fewer tokens than existing solutions in Claude and 1.75 times fewer in Codex.
Things to check before use
This system is suitable for developers who want to add personalized memory features to AI agents or chatbots. If you prefer local execution, all data is stored in the ./.supermemory directory, making backup and migration easy. You can freely choose models from OpenAI, Anthropic, Gemini, Groq, and other OpenAI-compatible endpoints, while embeddings use the local Xenova/bge-base-en-v1.5 by default. The README does not explicitly mention the license type or the scope of enterprise support, so you should verify Enterprise options and specific terms through the Self-hosting section of the official documentation before commercial deployment.
