DEV Community

Cover image for We open-sourced Synap Core, a self-hosted memory engine for AI agents
Anish Yadav
Anish Yadav

Posted on

We open-sourced Synap Core, a self-hosted memory engine for AI agents

Hey DEV community!

At Maximem, we've been working on memory infrastructure for AI agents.

Our managed Synap product recently achieved 95.6% answer accuracy on LongMemEval_S across all 500 questions, using just 7.4K context tokens per question.

But we didn't want to keep all this work behind a managed product.

So today, we're open-sourcing Maximem Synap Core! πŸš€

GitHub Repository

Synap Core is a self-hosted memory engine for AI agents. It transforms conversations into structured memories and makes relevant context available across sessions.

Here's what it handles:

  • Memory Extraction: Extracts meaningful information from conversations.
  • Entity Resolution: Understands when different references point to the same person or entity.
  • Temporal Reasoning: Tracks changes and avoids relying on outdated information.
  • Memory Evolution: Identifies duplicates, enriches existing memories, and updates outdated facts.
  • Hybrid Retrieval: Combines vector and graph retrieval to find relevant context.

What's under the hood?

We're using Python, PostgreSQL, ChromaDB, Neo4j and Redis, with Docker Compose to bring everything together.

You can get started with a single docker compose up.

It's licensed under Apache 2.0, so you can self-host it, explore the code, and build on top of it.

We'd love your help!

We're actively building Synap Core and would love to have more developers involved.

Whether it's improving retrieval, fixing bugs, optimizing performance, building integrations, or simply trying it out and sharing feedback, every contribution helps.

If you're building AI agents or experimenting with agent memory, we'd love for you to check it out.

⭐ Star the repository: https://github.com/maximem-ai/maximem_synap_core

🌐 Explore Maximem: https://www.maximem.ai/

🀝 Contributions and PRs are always welcome!

Would also love to hear how you're approaching memory in your own AI agents.

Let's build AI agents that actually remember!

opensource #ai #python #showdev #help

Top comments (1)

Collapse
 
macro8n profile image
macro8n | Desktop Automation •

For workflows that touch a changing UI, it seems important to separate stable facts from short-lived state and make updates traceable; stale context can otherwise lead an agent into the wrong next step. Nice to see this work open-sourced.