Skip to content

Add memory-type grouping for LMEB tasks: episodic, dialogue, semantic, and procedural memory #4898

Description

@ItsukiFujii

Description of the feature

Feature request

Currently, the Long-Horizon Memory benchmark (LMEB) is shown as a single Retrieval benchmark on the MTEB leaderboard. While this is technically correct because all LMEB tasks are evaluated as evidence retrieval tasks, it loses an important part of the benchmark structure.

LMEB is designed to evaluate long-horizon memory retrieval across four memory types:

  • Episodic memory
  • Dialogue memory
  • Semantic memory
  • Procedural memory

https://mteb-leaderboard.hf.space/benchmark/LMEB

However, the current MTEB page only reports the benchmark under Retrieval, with `Task Types = 1

This reduces the diagnostic value of the benchmark, especially for users who want to compare embedding models for long-term memory, agent memory, or memory-augmented RAG applications.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestleaderboardissues related to the leaderboard

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions