Introducing Aperture Nexus: Open Source Memory for AI Agents, Built on ApertureDB

If your agents work with images, documents, video, or structured records, not just chat logs, you've likely hit the same wall we did. Most memory tools treat everything as text chunks in a vector store. An agent walks into a repeat problem like it's the first time, every time, because the memory underneath it was never built to hold what the agent actually saw.

Aperture Nexus is our answer. It's built directly on ApertureDB, the unified graph-vector-multimodal database already running in production, so multimodal facts stay first-class instead of being reduced to captions or extracted text. Context, who, what, when, why, and how, is stamped on every commit. Every commit traces back to its original source, the image, the document, the actual session, today; lineage surviving across updates is a v2 item. Knowledge and Memory both live together in ApertureDB.

Most agents already have their own reasoning, Claude knows what's worth persisting as well as anyone does. This version of Nexus is built around that: a place for whatever reasoning is already happening to commit, retrieve, and build on, rather than introducing a competing reasoning layer of its own. Future versions will build on that foundation based on what we learn across different verticals.

This didn't come from a whiteboard. We audited over 20 existing memory frameworks to see what actually holds up, and shaped this against what we've learned deploying ApertureDB across organizations at very different stages of AI adoption. The thinking behind it: Introducing Aperture Nexus: Multimodal Memory and where existing memory tools fall short. We also started a podcast on this exact question, The Cognitive Layer: Episode 1, with a Netflix engineering lead on trustworthy AI memory.

v0.1, live now: commit, search, connect, authenticate.

What's next, public in the repo's discussions: Memory.update() with lineage, query-time visibility enforcement, GraphRAG neighborhood search, hybrid search, recency weighting, an MCP server.

It's open source, MIT licensed.

Try it: https://github.com/aperture-data/aperture-nexus
Docs: https://aperture-data.github.io/aperture-nexus/

Try it, and if it solves something you're dealing with, star it or send it to someone hitting the same wall.

Do share if something breaks when you use it or what else you would wish Nexus would support.

Regards,

Vishakha