AIBrain launches SelRoute, local memory

The useful part here is not “memory” as a buzzword, it’s the routing split: different query types need different retrieval paths, and most memory stacks still flatten that distinction.

• –SelRoute’s routing model is the right abstraction for conversational memory: factual, temporal, multi-hop, and summary queries should not all hit the same retriever.
• –The local-first setup matters for Ollama users because the memory layer stays on-device, with no GPU-heavy inference path just to fetch context.
• –The benchmark story is credible enough to matter: the arXiv paper reports 0.800 Recall@5 on LongMemEval_M and 62,000+ instances across additional benchmarks, but that is still a retrieval benchmark, not end-user product proof.
• –The paper also surfaces a real caveat: reasoning-heavy retrieval remains weak, so this looks like a strong systems improvement rather than a universal memory fix.
• –MCP support is the distribution win here, since it lets the system plug into other clients instead of trapping the memory layer inside one app.