Source: Mem0, the token-efficient memory algorithm (Mem0, 2026)

Mem0's April 2026 redesign: single-pass ADD-only extraction, entity linking and fused semantic/keyword/entity retrieval, with vendor-reported LoCoMo 92.5 and LongMemEval 94.4 at about 7k tokens per query.

Contents

Summary

Mem0 replaced the two-pass extract-then-reconcile pipeline from its 2025 paper with a single extraction call that only adds memories. Contradicted facts are kept alongside newer ones instead of being overwritten, and retrieval fuses several signals. All numbers are self-reported.

Key claims

  • Extraction: one LLM call instead of two; the ADD/UPDATE/DELETE reconciliation step is removed because it was slow and could erase useful information. Extraction latency roughly halves.
  • Facts stated by the agent (bookings, recommendations) are stored with the same weight as user facts.
  • Entity linking: proper nouns, quoted text and compound noun phrases are embedded into a separate lookup layer; memories sharing an entity with the query get a ranking boost.
  • Retrieval runs semantic, keyword (with verb normalisation) and entity matching in parallel and fuses them.
  • No dedicated temporal module; temporal reasoning relies on keeping history under ADD-only.
  • Results, old to new algorithm: LoCoMo 71.4 to 92.5; LongMemEval 67.8 to 94.4; BEAM 64.1 (1M tokens) and 48.6 (10M). About ±1 point judge noise. Average about 6.7k–7.0k retrieved tokens per query versus 25k+ for full context.
  • The post concedes LoCoMo and LongMemEval can be inflated by aggressive retrieval, big windows or frontier models, and calls BEAM the most relevant. Answer and judge models are not named.
  • Available in Platform and open source, but scores are from the Platform, which has proprietary optimisations; OSS users should expect only directionally similar gains. The eval harness is open-sourced.

Feeds

Mem0

Suggest changes →