The memory layer that never calls an LLM: what that buys, and what it costs

The article compares RE-call, a no-LLM memory layer that stores raw turns in Postgres, with Mem0, which uses an LLM to distill memories at write time.

On the BEAM benchmark, Mem0 does better in some areas, especially temporal reasoning, because compressed memories are easier to retrieve and disambiguate than scattered conversation logs.

The trade-off is cost and control: Mem0 adds token spend, slower ingest, and external data exposure, while RE-call avoids model costs and keeps memory local.

The author notes reranking likely matters too, but says the right choice depends on whether you prioritize accuracy or low-cost local storage.