HackerNews中文版

嗨，HN，我们正在构建Engram，一个具有持久记忆的AI助手，能够在不同会话之间有效工作。与ChatGPT/Claude不同的是，后者在每次聊天后会忘记一切，而Engram能够自动提取和索引事实、偏好和上下文。核心问题：大型语言模型（LLMs）在短期记忆方面表现出色，但在长期记忆方面几乎没有保留。如果你在一月份告诉Claude一个项目，它到三月份未必会记得，除非你把整个对话重新粘贴回来。我们的解决方案： - 使用14个因素的重要性评分算法自动提取记忆（目标/承诺的优先级高于随意事实） - 通过pgvector和OpenAI嵌入进行语义检索 - 认知画像，从写作样本中学习你的沟通风格 - 多供应商路由（例如，免费层使用Llama 3，优质层使用Gemini等）技术栈：React + Supabase（PostgreSQL + pgvector），全程使用TypeScript。我们构建了一个抽象层，处理供应商故障转移和速率限制。当前状态：处于测试阶段，约有300名来自宾夕法尼亚大学的用户（我们获得了沃顿创新基金建设奖）。第30天的留存率约为60%，这表明我们的记忆系统确实有用。试试吧：engramartificial.com 我之所以构建这个，是因为我对我的AI“助手”无法记住我昨天告诉它的事情感到沮丧。希望听到其他人的反馈，看看这是否引起共鸣，或者我是否在解决一个并不存在的问题。

查看原文

Hi HN,We are building Engram, an AI assistant with persistent memory that actually works across sessions. Unlike ChatGPT/Claude, which forget everything after each chat, Engram extracts and indexes facts, preferences, and context automatically.The core problem: LLMs have great short-term memory but zero long-term recall. If you tell Claude about a project in January, it won't necessarily remember in March unless you paste the whole conversation back in.Our approach: Automatic memory extraction using a 14-factor importance scoring algorithm (goals/commitments ranked higher than casual facts) Semantic retrieval via pgvector + OpenAI embeddings Cognitive profiling that learns your communication style from writing samples Multi-provider routing (Llama 3 for free tier, Gemini for premium to name a few)Technical stack: React + Supabase (PostgreSQL + pgvector), TypeScript throughout. We built an abstraction layer that handles provider failover and rate limiting.Current status: Beta with ~300 users from UPenn (we won the Wharton Innovation Fund Build award). Day-30 retention is ~60%, which suggests the memory system is actually useful.Try it: engramartificial.comBuilt this because I was frustrated that my AI "assistant" couldn't remember what I told it yesterday. Would love to hear if this resonates with others or if I'm solving a non-problem.

具备智能记忆的人工智能助手