# recall, compared

recall beside every memory system it has been measured against: one page each, with the shared benchmark, cost, latency, privacy and who should pick which.

- [recall vs Mem0](https://rererecall.com/compare/recall-vs-mem0/index.md): Memory API for apps and agents: an LLM extracts facts from each message into a vector store, hosted or self-hosted; Apache 2.0 with platform-only features. Where both were measured: SWE-ContextBench, past fix in the top 3, Mem0 39.4%, recall 76.8%.
- [recall vs Zep](https://rererecall.com/compare/recall-vs-zep/index.md): Context layer for enterprise agents: an LLM pipeline turns every message into a bi-temporal knowledge graph; Zep Cloud, with Graphiti (Apache 2.0) underneath. Where both were measured: LongMemEval_S, answers correct, Zep 90.2% (site) · 71.2% (paper), recall 88.4%.
- [recall vs Supermemory](https://rererecall.com/compare/recall-vs-supermemory/index.md): Hosted memory API with coding-agent plugins: its own model extracts and consolidates into a vector-graph store; MIT SDKs over a closed server, credits-based plans. Where both were measured: SWE-ContextBench, past fix in the top 15, Supermemory 59.6%, recall 97.0%.
- [recall vs LangMem](https://rererecall.com/compare/recall-vs-langmem/index.md): LangChain's memory SDK for LangGraph agents: memory tools and a background manager that prompt an LLM to write memories into a store you run; MIT, pre-1.0. Where both were measured: SWE-ContextBench, past fix in the top 10, LangMem 73.3%, recall 92.9%.
- [recall vs Mastra](https://rererecall.com/compare/recall-vs-mastra/index.md): TypeScript agent framework whose observational memory has an observer model compress the conversation and a reflector condense it; Apache 2.0 core, a hosted platform. Where both were measured: LongMemEval_S, answers correct, Mastra 94.9%, recall 88.4%.
- [recall vs Honcho](https://rererecall.com/compare/recall-vs-honcho/index.md): User-modelling memory for apps: a background deriver reasons over every message into a representation of each person; AGPL, hosted with Plastic Labs' own models. Where both were measured: LongMemEval_S, answers correct, Honcho 90.4%, recall 88.4%.
- [recall vs Hindsight](https://rererecall.com/compare/recall-vs-hindsight/index.md): Open-source memory engine with four memory networks: an LLM extracts on every retain, retrieval fuses dense, BM25, graph and temporal search with a reranker; MIT, with a cloud. Where both were measured: LongMemEval_S, answers correct, Hindsight 94.6% (board) · 91.4% (paper), recall 88.4%.
- [recall vs EverMemOS](https://rererecall.com/compare/recall-vs-evermemos/index.md): Open-source memory runtime, now EverOS: a model splits chats into memory cells and scenes and verifies retrievals at query time; Apache 2.0, a $25 cloud after beta. Where both were measured: LongMemEval_S, answers correct, EverMemOS 83.0%, recall 88.4%.
- [recall vs OpenViking](https://rererecall.com/compare/recall-vs-openviking/index.md): ByteDance's context database for agents: a filesystem-like tree of knowledge, memory and skills, each node summarised by a model; AGPL, with a hosted service in Beijing. Where both were measured: SWE-ContextBench, past fix in the top 3, OpenViking 51.5%, recall 76.8%.