Dash Research Vault

概述 | Overview

A multi-agent research vault implementing Evensong's three core findings:

Finding	Application	Mechanism
Memory Causation	Agent knowledge base accumulates over time	Dialogue extraction → Admin review → Write
Pressure-Triggered Self-Evolution	L1 self-drive criteria in each agent description	Cross-review in #internal
Recursive Contamination Control	Three-tier gate (Agent→Admin→Author)	Write requires review to prevent low-quality self-reinforcement

触发词 | Triggers

English: research vault, memory vault, research-vault, dash vault, vault skill
中文: 研究 vault, 记忆库, 研究库, vault 技能, 记忆 vault

目录结构 | Directory Structure

vault/
├── 00-overview/          ← Vault index + key figures
├── 01-agents/            ← Agent role definitions
├── 02-channels/          ← Communication channel types
├── 03-internal-only/     ← Privacy-controlled content
├── 04-memory/            ← Three-tier memory with decay
│   ├── public/           ← All agents readable
│   ├── restricted/        ← Admin-only access
│   └── .meta/            ← Decay metadata
└── docs/                  ← Research documents

核心机制 | Core Mechanisms

衰退管理 | Decay Management

Based on Ebbinghaus forgetting curve, entries decay by difficulty:

Difficulty	Retention Rate	Use Case
4.0	Very slow	Core client/industry facts
2.0	Medium	Technical knowledge
0.5	Fast	Dialogue extracts (unless cited)

Decay formula: retention = e^(-time / (difficulty * half_life))

三层权限 | Three-Tier Access

• public/: All agents can read and write
• restricted/: Admin-only read, write via review gate
• internal-only/: No agent access, Researcher-only

使用示例 | Usage Examples

# Initialize a new vault
dash-research-vault init --name "My Research"

# Run weekly decay maintenance
dash-research-vault decay --tier public

# Scan and sync new findings
dash-research-vault loop --tier restricted

关键图表 | Key Figures

Figure 1: Swarm Taxonomy

Figure 2: Behavioral Heatmap

Figure 3: Memory Causation Chain

Figure 4: L2 Pressure Self-Evolution