智能记忆 AI 助手
1 分•作者: jwallace•6 个月前
Hi HN，
我们正在构建 Engram，一个具有持久记忆的 AI 助手，可以在不同会话中真正发挥作用。与每次聊天后都会忘记所有内容的 ChatGPT/Claude 不同，Engram 会自动提取并索引事实、偏好和上下文。
核心问题：LLM 具有出色的短期记忆，但长期记忆为零。如果你在一月份告诉 Claude 关于一个项目的事情，它在三月份不一定会记住，除非你把整个对话重新粘贴进去。
我们的方法：使用 14 因素重要性评分算法进行自动记忆提取（目标/承诺的排名高于随意的事实）
通过 pgvector + OpenAI 嵌入进行语义检索
通过写作样本学习你的沟通风格的认知分析
多提供商路由（Llama 3 用于免费层，Gemini 用于高级层，仅举几例）
技术栈：React + Supabase (PostgreSQL + pgvector)，全程使用 TypeScript。我们构建了一个抽象层，用于处理提供商故障转移和速率限制。
当前状态：Beta 测试版，约有 300 名来自宾夕法尼亚大学的用户（我们赢得了沃顿创新基金建设奖）。第 30 天的留存率为约 60%，这表明记忆系统确实有用。
试用：engramartificial.com
构建这个是因为我感到很沮丧，我的 AI “助手” 无法记住我昨天告诉它的事情。很想知道这是否引起了其他人的共鸣，或者我是否在解决一个非问题。
查看原文
Hi HN,<p>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.<p>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.<p>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)<p>Technical stack: React + Supabase (PostgreSQL + pgvector), TypeScript throughout. We built an abstraction layer that handles provider failover and rate limiting.<p>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.<p>Try it: engramartificial.com<p>Built 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.