I’ve been building Synrix, a local-first memory engine for AI systems.

The idea is simple: instead of pushing memory into vector databases or cloud services, Synrix runs entirely on your machine and focuses on predictable, targeted recall rather than global similarity scans.

Architecturally it’s different from typical vector DB approaches:

Queries scale with matching results (O(k)), not total dataset size

Runs fully local (no network calls, no cloud dependency)

One binary, tier controlled by signed key (SDK is MIT, engine is proprietary for scale)

Designed for structured + semantic memory (agents, RAG, task stores, etc)

On local datasets (~25k–100k nodes) we’re seeing microsecond-scale prefix lookups on commodity hardware. We haven’t published formal benchmarks yet, but plan to add reproducible tests soon.

I’m especially interested in feedback from people who’ve built agent memory systems, RAG pipelines, or dealt with scaling vector databases.

Questions I’d love input on:

Do you think local-first memory makes sense for agents, or does cloud still win?

Have vector DBs been working well for you at scale?

What would you want to see in benchmarks?

Repo here: https://github.com/RYJOX-Technologies/Synrix-Memory-Engine

Happy to answer anything.