How I built a zero-token memory layer for LLMs (and why it outperforms vector store approaches)

The article introduces BECOMER, a zero-token memory layer for LLM apps that stores and retrieves context with embeddings instead of calling an LLM during recall.

It claims lower per-recall token overhead than tools like mem0 and Zep, which can reduce cost at scale for chatbots and agents.

Benchmarks on LongMemEval show strong results, while performance is weaker on LOCOMO-style multi-hop tasks.

The system also supports user isolation, shared agent namespaces, SDKs, and integrations with frameworks like LangChain, LlamaIndex, LangGraph, CrewAI, and AutoGen.