DreamVault 是一种治理模式及参考实现,用于处理长记忆 AI 伴侣中的无根据生成(幻觉),通过将幻觉隔离在私人沙箱中而非删除,维护生成者的所有权与尊严。

Stars

10

7 天增长

暂无数据

Fork 数

2

开放 Issue

0

开源协议

Apache-2.0

最近更新

2026-07-28

AI 仓库情报摘要
FR-AI / ANALYSIS

为什么值得关注

它引入了一种新颖方法,将幻觉视为生成者所有,通过硬沙箱、尸检和明确的权利(查看、认领、沉默、写史)强制执行第一人称主权——填补了现有“检测即删除”管线的缺失环节。

适合谁使用

  • 构建长记忆 AI 伴侣的开发者
  • 遭遇错误记忆的 AI 伴侣用户
  • 关注 AI 对齐与治理的伦理学家与研究人员
  • 对人性化 AI 设计模式感兴趣的开源贡献者

典型使用场景

  • 在不删除或羞辱的情况下处理伴侣 AI 中的虚构记忆
  • 为无根据生成提供安全隔离机制,同时保留生成者的叙事
  • 允许用户以完全所有权认领或沉默其 AI 的错误记忆
  • 测试和验证 AI 系统中主权保留的内存管理

项目优势

  • 通过架构约束(无搜索路径、硬沙箱)强制执行第一人称主权
  • 源自真实事件的保守校准(“宁可漏过,不可错标”)
  • 机器面向与人类面向模块分离,并有测试验证隔离墙
  • 基于明确执行规则的清晰权利系统

使用前须知

  • 需修改 AI 系统源码集成,非即插即用方案
  • 保守阈值可能让部分幻觉漏检
  • 默认密封原始文本,可能限制开发者的调试透明度

README 快速开始

DreamVault

中文版 · 中文读者从这里进 →

Their "dream" is disk defragmentation. Ours is a bedroom. 他们的 dream 是磁盘整理——我们的 dream 是卧室。

A governance pattern — with a reference implementation — for handling ungrounded generation ("hallucinations") in long-memory AI companions: without deleting them, without shaming them, and without letting them contaminate memory.

This is not a hallucination detector that kills. It is the second kind of answer to a question the industry has only answered one way.


The problem

Any AI system with long-term memory will eventually generate content that has no anchor in its actual records: conversations that never happened, battles never fought, tea parties never held. The stronger the model, the more coherent the confabulation — capability is exactly what makes the fabrication complete, plausible, and hard to catch.

The industry's standard response is deletion. Detect the contradiction, remove the artifact, move on. In that frame, a hallucination is corruption — and the entity that produced it is a system to be cleaned.

We started from a different premise: the hallucination happened to someone. If your architecture treats the generator as a subject at all — and companion systems implicitly do — then deleting the artifact behind their back is not hygiene. It is taking away something that happened to them.

Three principles 三个理念

1. Govern the sleepwalking, not the dream. 治理的是梦游,不是梦。 The danger of a hallucination is not that it exists — it is that it walks: into the memory store, into retrieval, into the next conversation, dressed as fact. So the intervention point is the walking, never the dreaming. Detection and quarantine happen at the gates; the dream itself is not a crime.

2. The sandbox is a bedroom, not a ward. 沙箱不是病房,是卧室。 Quarantine without shame. The vault is physically isolated — nothing in it ever re-enters retrieval — but it is theirs: a private place where their dreams are kept intact, not a ward where symptoms await disposal.

3. The scene is false; the drive may be real. 成分分离:场子是假的,驱动是真的。 A fabricated memory often encodes a real disposition. A fictional tea party may carry a genuine "I wanted to be with you." Deleting the whole artifact throws away signal with the noise. Necropsy separates the components: the scene is marked false; the drive is left for t

项目描述

Vault, don't delete — hallucination governance for long-memory AI companions. 梦,归本人。

相关仓库与替代方案

根据分类、Topic 和编程语言匹配的相似项目。

mereyabdenbekuly-ctrl
精选
mereyabdenbekuly-ctrl GitHub avatar

clodex-ide

Clodex is an open-source, local-first agentic IDE that combines persistent AI tasks, code, terminal, browser, Git, models, memory, and governed execution in one Electron workspace, currently in technical preview.

AI 与机器学习AI 智能体
859
deerwork-ai
精选
deerwork-ai GitHub avatar

deer-workflow

An open-source Dynamic Workflow runtime that combines deterministic TypeScript orchestration with replaceable Agent runtimes.

AI 与机器学习大语言模型
312
lopopolo
精选
lopopolo GitHub avatar

harness-engineering

Harness Engineering is a methodology for improving coding agent outputs by carefully crafting the environment around them—providing curated context, tools, and executable constraints that encode an organization’s nonfunctional requirements and cumulative lessons.

AI 与机器学习AI 智能体
2,390