AI UNDERDOGSDAILY PICK
AI UNDERDOGS
AI 记忆，该激活而非检索
Memory Should Activate, Not Retrieve
CrypticCortex/agam
大多数 agent 记忆方案
本质就是 RAG 加个壳
但你回忆事情的时候不是在脑子里做搜索——是某个场景把记忆激活了
Agam 就是按这个思路来的
Most agent memory projects?
RAG with extra steps. But you don't search
your brain every morning — context activates memories
Agam builds on that idea
关联式记忆图谱
Associative Memory Graph
Agam 用知识图谱存记忆
每个节点和别的节点有关联
当新的上下文进来
相关的记忆节点被自动激活——更像人脑的联想记忆
不是去数据库里翻
Agam stores memories in a knowledge graph. When
new context arrives, related nodes activate automatically —
closer to how associative recall actually works than
hitting a vector database
★ SIGNAL 1
记忆会过期、会合并
Memories Decay and Merge
这个项目的 prompts 目录里有一整套记忆生命周期管理——过时的记忆要标记
损坏的要修复、碎片要合并
老化的要升级
大多数项目写进去就不管了
这个真的在想记忆怎么变老
The prompts directory has a full lifecycle system
— obsolete, repair, digest, upgrade. Most projects store
and forget. This one actually thinks about memory
aging
★ SIGNAL 2
「记忆」两个字，想清楚了
They Thought About What Memory Means
HN 帖子里作者直接画了条线：这不是给向量搜索穿个马甲
光是这个区分——激活 vs 检索——就说明作者不是在做 demo
是在想底层模型到底该怎么设计
The author drew a hard line on HN
this isn't vector search in disguise. That distinction
alone — activation vs retrieval — shows they
actually thought about what memory means at a
model level
目前 Agam 主要面向 Claude Code 和 Codex 这类编码 agent
项目还在早期阶段
但光是把「记忆该有生命周期」这个想法做出来
就比大多数存储方案高了一个维度
Agam targets coding agents like Claude Code and
Codex right now — still early. But even
just implementing the idea that memories need a
lifecycle puts it a dimension above plain storage
solutions
AI UNDERDOGS
开源，给你的 agent 装个真脑子
Open source — give your agent a real brain
CrypticCortex/agam
关注 · 每天发现更多 AI 神作
github.com/CrypticCortex/agam