Memory Stopped Being Retrieval
I said in 2025 that retrieval strategy beats storage and that similarity alone would not carry it. A year later the benchmarks and the tooling agree, and they put a number on what it actually takes.
I wrote a note called Vector Memory for Agents back in September 2025: storage is cheap, retrieval strategy is what matters, and you need semantic similarity plus temporal relevance and interaction frequency. That was a bet at the time. The benchmarks have caught up since.
The 2026 numbers back it. Mem0's April report puts the gains on two changes: multi-signal retrieval, semantic similarity plus keyword and entity matching, and single-pass extraction that treats agent-generated facts as seriously as user statements. +29.6 points on temporal reasoning, +23.1 on multi-hop. It still degrades with scale, 64.1 on BEAM at one million tokens falling to 48.6 at ten. Pinecone shipped Nexus into public preview on July 1st with the same conclusion baked into the architecture: compile the context upstream, hand the agent structured knowledge instead of raw documents.
So the direction held and the bar moved. Similarity alone was never enough - it takes structure, recency, and provenance to decide what stays. Steering coding agents taught me the same lesson from the other side: the model is rarely the bottleneck, the context I hand it is.