Skip to main content
Back

Mem0 Rewrites Its Memory Algorithm, Hitting 94.4% Accuracy at 4x Lower Token Cost

Mem0 ships a rewritten memory algorithm with +27pt accuracy gains, single-pass extraction, entity linking, and temporal reasoning — at 3-4x lower token cost

  • Mem0 shipped a new memory algorithm with 92.5% on LoCoMo (+21 pts) and 94.4% on LongMemEval (+27 pts), at ~7K tokens per query vs 25K+ for full-context.
  • Architecture switches to single-pass ADD-only extraction — one LLM call per turn, no UPDATE/DELETE operations, with entity linking and multi-signal retrieval.
  • Biggest gains are on temporal queries (+29 pts) and multi-hop reasoning (+25 pts), the two hardest categories for prior memory systems.
  • Agent-native signup flow lets AI agents provision their own Mem0 API key via CLI with no human dashboard interaction required.
  • Free tier offers 10K memories/month; paid plans from $19/mo (Starter) to $249/mo (Pro); graph memory locked to Pro and above.
  • Open-source (Apache 2.0) with 62.8K GitHub stars; benchmarks and paper are publicly reproducible.

Mem0 (pronounced "mem-zero") is the leading open-source memory layer for AI agents, sitting at 62,800 GitHub stars since its YC S24 debut. The core idea: instead of stuffing entire conversation histories into every prompt, Mem0 extracts the facts worth keeping, stores them in a vector database, and retrieves only what's relevant at query time. The result is a stateful agent that remembers users without burning through tokens.

The project just shipped its most significant architectural overhaul: a fully rewritten memory algorithm that replaces the old extract-update-delete cycle with a cleaner, faster, and substantially more accurate approach.

The benchmark numbers

The new algorithm was tested across three standard memory benchmarks:

  • LoCoMo covers 1,540 questions spanning single-hop, multi-hop, open-domain, and temporal recall.
  • LongMemEval runs 500 questions across categories including knowledge updates and multi-session recall.
  • BEAM evaluates at 1M and 10M token scales.
BenchmarkOldNewTokens/queryLatency p50
LoCoMo71.492.57.0K0.88s
LongMemEval67.894.46.8K1.09s
BEAM (1M)—64.16.7K1.00s
BEAM (10M)—48.66.9K1.05s

The previous algorithm already ran at 3-4x lower token cost than full-context approaches, which routinely consume 25,000+ tokens per query. The new one keeps that efficiency while closing most of the accuracy gap. The two biggest gains are on temporal queries (+29.3 points) and multi-hop reasoning (+25.2 points), meaning the system now handles questions like "when did the user first mention X?" or "what led to the user's current decision?" far more reliably.

What changed under the hood

The old algorithm treated memory as a mutable store: when new information arrived, it would UPDATE or DELETE existing memories to maintain consistency. That required multiple LLM calls per conversation turn and introduced errors whenever the model misjudged whether a new fact superseded an old one.