ConvMemory: The Tiny AI That's Lowkey Changing Memory Retrieval

Ok wait because this is actually insane. There's a new player in the AI game and it's called ConvMemory. It’s like that quiet kid in class who's secretly a genius. We're talking about a tiny 3.6 million parameter reranker that's making waves conversational long-term memory retrieval. And yes, it's doing it with style.

Speed and Savings

Here's the tea: ConvMemory operates with 12-47 times lower latency than the BGE-large cross-encoder. Bestie, that's fast. But it doesn't stop there. While it's zooming past its competitors, it also runs on the cheap. Like 28 times cheaper than the mxbai-rerank-large-v1 on the Clean500 dataset. No cap. And even when it faces off with Stress1000 distractors, the latency is 117 times lower. Cha-ching.

The Real Deal

Now let's get real. ConvMemory isn't all about speed. It's also about efficiency. It uses this thing called cross-encoder teacher supervision over fused dense and lexical features. Sounds fancy, but basically, it's getting the job done without the frills. It doesn't try to exploit temporal structures, it just focuses on what works. That's a slay.

Why Should You Care?

No but seriously. Read that again. In a world where big tech loves to flex with massive models, ConvMemory is showing us that small can be mighty. It's like the David to Goliath's big, clunky models. And let's be honest, who doesn't love an underdog story?

Plus, the team behind ConvMemory released CCGE-LA, a conflict-aware candidate-set editor. It's not perfect, but it's making consistent progress on LoCoMo's supersession and stale/rescue slices. That's like finding hidden treasure in your backyard.

So, what's the catch? ConvMemory doesn't match the mxbai-rerank-large-v1 in absolute LoCoMo MRR. But do we care? It's doing its own thing, and it's doing it well. The way this protocol just ate. Iconic.