正在打开每日关注
正在打开每日关注
以单次追加、实体链接和多信号检索构建长期记忆，并用基准数据证明性能提升。
my opus 5 + gpt 5.6 bill went from $250/month to $18 - i was reexplaining my entire project to the agent every single morning everyone points you at mem0 - 62,000 stars, 3 years old, the "default answer". then there's a 4-month-old repo sitting right behind it with 58,000 stars the first line of its readme just says: it beats every other memory system on benchmarks. no hedging, no maybe 3 steps and your agent stops forgetting between sessions install --> load project once --> agent remembers everything --> zero context re-explanation it's collecting stars 8x faster than mem0. at this rate it takes the crown in 2 weeks no vector db rebuild, no rag pipeline swap, no api rewrite - just replace one library so which one did you install - the "default answer" or the one that actually beats it on benchmarks bookmark this before mem0's team panics - one library swap replaces $250/month in context reload spend
是什么 这是一条以“Agent 不再遗忘项目上下文”为卖点的记忆库推荐。推文把 Opus 5/GPT-5.6 的账单下降、Mem0 星标规模与另一方案的增长速度串成替代论证；费用、星标和“击败所有系统”是作者主张，截图不能独立证实。画面中的 GitHub README 对应 Mem0，定位为面向 AI Agent 的通用记忆层。
核心设计
README 展示 single-pass、ADD-only 提取：一次 LLM 调用写入事实，不做 UPDATE/DELETE；配合 Agent 生成事实、实体链接、多信号检索和时间推理增强召回。LoCoMo、LongMemEval、BEAM 表格比较新旧分数、token 与 p50 延迟；代码以 memory.add 接入对话调用。
关键洞察 可借鉴的边界是把抽取、索引、检索集中到记忆层，应用侧保留模型调用，从而支持个性化聊天、编码 Agent、LangGraph/CrewAI 等集成，减少跨会话上下文重载。README 还列出浏览器扩展与 ChatGPT 记忆入口。
边界与风险 ADD-only 可能造成事实累积、冲突和过期；删除、隐私隔离、租户安全与合规不会因“换一个库”自动解决。基准收益依赖数据集、模型栈和召回配置，生产选型仍应实测 token、延迟与迁移成本。