Show HN: MemorySync – Memory Layer for AI agents, with open-source integrations

Rafay, founder of MemorySync, launched MemorySync, a managed memory API that lets AI agents save and recall user preferences and past decisions across sessions, with open-source MIT-licensed integrations for Claude Code, Cursor, n8n, Dify, Pipecat, and LiveKit. The service stores facts as nodes in a graph organized into hot, warm, and cold tiers, scores recall by semantic similarity, recency, and importance, and prices by saves and searches rather than tokens, with a free plan of 10k saves and 1k searches per month. The core memory engine remains closed source, while a benchmark repository is available at github.com/memorysyncio/memory-benchmarks.

Hey HN I'm Rafay, the founder of MemorySync https://memorysync.io https://memorysync.io . I built MemorySync to help AI agents remember users, their preferences, and past decisions through a single API. Here's a quick demo: https://youtu.be/ylfoN 5oXTI https://youtu.be/ylfoN 5oXTI You can try it without signing up at https://memorysync.io/try https://memorysync.io/try . Save a few memories, refresh the page, and ask the agent to recall them. If you've built an AI agent that people come back to, you may have run into this problem. A user tells the agent on Monday that they prefer short answers and have moved to another city. By Friday, the agent asks them the same questions again. One solution is to send the entire chat history with every prompt. But that increases token costs and latency, and the agent can still miss important details. I ran into this while building coding agents and live voice agents. Even with the full chat history, agents could contradict themselves when something changed. I also kept copying the same memory logic into different integrations. I wanted a way to handle memory once and use it across my agents without rebuilding the same logic for each one. That's what led me to build MemorySync, a single API for saving and recalling memories that works with existing agents, including Pipecat and LiveKit. When something changes, MemorySync updates the relevant information, whether it's a user's city, role, or preferences. A user record is created when the first memory is saved, so information carries across sessions and returning users don't have to repeat themselves. Here's how it works: Adding memories: Your agent sends conversations or facts to MemorySync. A model identifies information specific to the user that's likely to be useful later, such as preferences, plans, and relationships. Greetings and filler are filtered out, and duplicate facts are stored only once. Storing memories: Instead of raw chat logs, MemorySync stores memories in a graph. Each node represents a fact, and connections show which facts are related, support each other, or contradict each other. When a current value changes, such as a user's city, role, or theme, the newer fact remains searchable and the older one is deleted. Memory tiers: Memories are organized into hot, warm, and cold tiers. Frequently used context stays quick to access, while older context becomes cheaper to store. Memories that aren't used often gradually move to lower tiers. Retrieving memories: When an agent needs context, it asks MemorySync. Memories are scored by semantic similarity, recency, and importance. Preferences rely more on importance, while questions about what's currently true rely more on recency. MemorySync returns the best matches. Recall is designed around a latency budget, so voice agents don't have to wait long for memory. With Pipecat and LiveKit, callers preferences and past decisions can be available before a call starts, helping the agent recognize returning users. Instead of sending the full conversation history with every turn, the agent gets only the context it needs. Pricing is based on saves and searches, not tokens. The core memory engine is a closed-source managed service. The integrations are open source under the MIT license and include Claude Code, Cursor, n8n, Dify, Pipecat, and LiveKit. There's also a benchmark repository you can run yourself: https://github.com/memorysyncio/memory-benchmarks https://github.com/memorysyncio/memory-benchmarks You can find the documentation at https://docs.memorysync.io https://docs.memorysync.io . The free plan includes 10k saves and 1k searches per month, with no credit card required. I'd love to hear how you're handling memory in your own agents. Have you run into similar problems? If you've solved them differently, I'd like to hear how. And if you haven't needed a separate memory API, what would make one useful to you? Give it a try and let me know what you think. I'm happy to answer questions, hear different approaches, and learn what I should improve next. Comments URL: https://news.ycombinator.com/item?id=50119814 https://news.ycombinator.com/item?id=50119814 Points: 1 Comments: 0