Memori Claims 81.95% LoCoMo Accuracy at 4.97% of Full-Context Tokens
MemoriLabs shipped Memori, an LLM-agnostic structured agent memory layer that claims 81.95% accuracy on the LoCoMo long-context memory benchmark at approximately 1,294 tokens per query — about 5% of full-context footprint, 67% smaller prompts than Zep, and more than 20 times cheaper than full-context retrieval. Memori uses an attribution model that tracks entities, processes, and sessions, exposes an MCP endpoint, and integrates with Pydantic AI, LangChain, and Agno.
Why It Matters
The LoCoMo numbers are the most concrete validation in this week's wave of agent-memory releases. If the benchmark holds under independent review, Memori offers a viable path to efficient long-context memory without the cost or latency of full-context retrieval — relevant for any production agentic system where conversation history needs to persist across sessions.