# Mem0 Algorithm Update Hits 91.6 on LoCoMo, 94.8 on LongMemEval > Mem0, the open-source memory layer for AI agents maintained by mem0ai, has released a benchmarked algorithm update that scores 91.6 on LoCoMo and 94.8 on LongMemEval, gains of 20 and 27 points respectively over its prior version, using single-pass retrieval with no agentic loops. - Canonical URL: https://agentry.press/news/mem0-algorithm-update-hits-91-6-on-locomo-94-8-on-longmemeval/ - Type: News - Published: 2026-06-04 - By: agentry - Tags: mem0, agent-memory, benchmarks, retrieval, open-source, ai-agents --- ## What Mem0 Is Mem0 (pronounced 'mem-zero') is an open-source, persistent memory layer built for AI assistants and autonomous agents [1]. Maintained by mem0ai, the library stores and retrieves user preferences, session context, and agent state across interactions, enabling personalized behavior over time. Primary deployment contexts include customer support chatbots that need to recall past tickets, AI assistants requiring consistent context across conversations, healthcare applications tracking patient history, and autonomous systems that accumulate operational knowledge [1]. ## Algorithm Changes in the New Release The updated algorithm introduces five discrete technical changes, each targeting a specific weakness in how prior versions managed memory [1]. The first change is single-pass ADD-only extraction. Each memory operation now requires one LLM call, and the system no longer issues UPDATE or DELETE operations. Memories accumulate rather than overwrite, which eliminates the multi-step agentic loops that previously added latency and complexity. The second change elevates agent-generated facts to first-class status. When an agent confirms an action, that confirmation is stored with equal weight to user-provided information, closing a gap where agent-side context was previously treated as secondary [1]. Third, entity linking now extracts, embeds, and links named entities across the memory store. Those links are then used to boost retrieval relevance when a query involves a known entity. Fourth, the retrieval pipeline moves to a multi-signal architecture, combining semantic search, BM25 keyword matching, and entity matching. All three signals are scored in parallel and fused into a single ranked result [1]. Fifth, temporal reasoning adds time-aware ranking so the system can distinguish queries about current state, past events, and upcoming plans, returning the correctly dated memory instance in each case [1]. ## Benchmark Results The updated algorithm was evaluated on three benchmarks, all run on the same production-representative model stack [1]. On LoCoMo, Mem0 scores 91.6, a gain of 20 points over the previous algorithm version. On LongMemEval, the score reaches 94.8, a gain of 27 points overall, with assistant memory recall specifically improving by 53.6 points. On BEAM at 1 million tokens, described as a production-scale memory evaluation, the system scores 64.1 [1]. All benchmark runs use single-pass retrieval with no agentic loops, meaning the numbers reflect one-call performance rather than multi-step chain results [1]. ## How Retrieval Works The multi-signal retrieval pipeline is the architectural centerpiece of the release. When a query arrives, three matching strategies run in parallel: semantic search over embedded memory vectors, BM25 keyword matching against stored text, and entity matching against the linked entity graph [1]. The parallel scores are then fused into a single ordered result set before being returned to the calling agent or assistant. Temporal reasoning operates as an additional ranking layer on top of that fused result, ensuring that time-sensitive queries surface the appropriate dated instance rather than the most recently added or most semantically similar memory regardless of when it was recorded [1]. Because extraction is ADD-only and single-pass, the retrieval index grows monotonically. Operators should account for index size over long-running deployments, though the BEAM evaluation at 1 million tokens suggests the pipeline was tested at meaningful scale [1]. ## Who This Targets and How to Adopt It The release is aimed at teams running AI assistants and autonomous agents in production where memory continuity and recall accuracy are operational requirements [1]. The benchmark improvements on assistant memory recall, specifically the 53.6-point gain on LongMemEval, are most relevant to customer support and healthcare deployments where recalling prior interactions accurately affects outcome quality. Existing users can consult the migration guide published alongside the release for upgrade instructions [1]. The evaluation framework used to produce the benchmark numbers has been open-sourced, allowing operators to reproduce results against their own data and model configurations rather than relying solely on the published figures [1]. ## FAQ **Q. Does the ADD-only extraction model cause unbounded memory growth?** Because the new algorithm accumulates memories without issuing UPDATE or DELETE operations, the memory store grows over time [1]. The BEAM benchmark at 1 million tokens indicates the retrieval pipeline was evaluated at scale, but operators managing long-running deployments should plan for index growth. **Q. Can the benchmark numbers be reproduced independently?** Yes. The evaluation framework used to generate the LoCoMo, LongMemEval, and BEAM scores has been open-sourced, so teams can run the same evaluations against their own model stacks [1]. **Q. Is there a migration path for teams on the previous algorithm?** A migration guide is available for existing users upgrading to the new algorithm version [1]. The source repository points to this guide directly alongside the benchmark documentation. **Q. What retrieval latency trade-offs come with parallel multi-signal scoring?** The sources do not provide specific latency figures for the multi-signal pipeline. The published benchmarks focus on accuracy scores rather than throughput or response-time metrics [1]. **Q. Does the system require a managed service, or can it run self-hosted?** Mem0 offers both an intuitive API with cross-platform SDKs and a fully managed service option, giving teams the choice between self-hosted and hosted deployments [1]. ## Key takeaways - Mem0's updated algorithm scores 91.6 on LoCoMo (+20 points) and 94.8 on LongMemEval (+27 points) compared to the previous version, using single-pass retrieval with no agentic loops [1]. - Five specific changes drive the gains: ADD-only extraction, first-class agent-generated facts, entity linking, multi-signal retrieval fusion, and temporal reasoning [1]. - The multi-signal retrieval pipeline runs semantic search, BM25 keyword matching, and entity matching in parallel, then fuses the results into a single ranked output [1]. - The evaluation framework is open-sourced, enabling independent reproduction of benchmark results [1]. - A migration guide is available for teams upgrading from the prior algorithm version [1]. ## Sources 1. https://github.com/mem0ai/mem0