Dash Research Vault

The knowledge storage algorithms and multi-agent harness that let one model
remember across sessions — and stop building the same thing twice.

🇺🇸 English · 🇨🇳 中文

▶ Watch the 2-minute film  ·  Live demo hub  ·  Read the paper

---

What this is

A vault template for multi-agent systems that need persistent memory. It encodes three findings from the Evensong benchmark series:

Finding	What it gives you	How it works
Memory Causation	A knowledge base that grows from real conversations, not curated docs	Dialogue extraction → Admin review → Write
Pressure-Triggered Self-Evolution	Agents that adapt their own strategy under load instead of breaking	L1 self-drive criteria + cross-review in #internal
Recursive Contamination Control	A guarantee that agents can't poison each other's memory	Three-tier write gate: Agent → Admin → Researcher

The vault is the substrate that makes those findings reproducible. Drop it in next to any multi-agent stack and you have a memory that survives session boundaries.

---

Directory structure

dash-research-vault/
├── 00-overview/         ← Vault index + 4 key figures
├── 01-agents/           ← 5 agent role definitions (Ops, Admin, Fin, Industry, Observer)
├── 02-channels/         ← 6 channel types (#internal, #ops, #admin, ...)
├── 03-internal-only/    ← Researcher-only strategy. No agent ever reads here.
├── 04-memory/           ← Three-tier memory + Ebbinghaus decay
│   ├── public/          ← All agents readable (client / dialogue / industry)
│   ├── restricted/      ← Admin-only (evolution / relationship / risk)
│   └── .meta/           ← Decay configuration
├── packages/
│   └── research-vault-mcp/ ← Installable MCP server for vault search/status tools
└── docs/                ← Research papers (EN + ZH)

---

Vault tiers

Tier	Read access	Write gate	Decay (default difficulty)
public/	All agents	Free	Fast — 0.5 (36h half-life)
restricted/	Admin only	Review required	Medium — 2.0 (144h half-life)
internal-only/	Researcher only	Direct	Slow — 4.0 (288h half-life)

The tier you pick determines how fast a memory fades and who can write to it. The decay is per-document, not global — so one important client fact at difficulty=4.0 outlives a thousand transient dialogue fragments at 0.5.

---

Decay algorithm

Based on the Ebbinghaus forgetting curve with per-document difficulty weighting:

retention(t) = e^(-t / (difficulty × half_life))

Default half-life: 72 hours. Multiply by difficulty to get the effective half-life per document.

Difficulty	Effective half-life	Use it for
4.0 (core)	288h ≈ 12 days	Client facts, industry positioning, hard-won lessons
2.0 (technical)	144h ≈ 6 days	Technical knowledge, runbook entries, system invariants
0.5 (dialogue)	36h	Conversation extracts, transient context, scratchpad

When a document's retention(t) drops below the read threshold, it stops surfacing in agent context — but stays on disk. Restore it by accessing it (touches last_accessed, resets the curve).

---

Key figures

All four figures are from the Evensong R012-E benchmark paper (docs/).

Figure	Topic
	Swarm taxonomy — how the 5 agents cluster by behavior under L0/L1/L2 pressure
	Behavioral heatmap — per-task per-tier response signatures
	Memory causation chain — dialogue → strategy recall → architecture decision
	L2 pressure timeline — when self-evolution triggers fire and what they change

---

Research papers

Paper	Language	File
When Agents Remember, They Stop Building	English	evensong-paper-en.pdf
当 Agent 记得，他们就停止重建	中文	evensong-paper-zh.pdf

Hengyuan Zhu (University of South Carolina). April 2026. arXiv preprint.

Citation (BibTeX)

@misc{zhu2026evensong,
  author       = {Hengyuan Zhu},
  title        = {When Agents Remember, They Stop Building: How Persistent Memory Alters AI Agent Engineering Strategy},
  year         = {2026},
  month        = {April},
  howpublished = {Preprint},
  institution  = {University of South Carolina},
  url          = {https://github.com/Fearvox/dash-research-vault}
}

---

Setup

# Clone
git clone https://github.com/Fearvox/dash-research-vault.git
cd dash-research-vault

# Adapt to your team
#  1. Edit roles in        01-agents/
#  2. Rename channels in   02-channels/
#  3. Tune decay in        04-memory/.meta/decay-config.md
#  4. Replace example data with your own (use placeholders for sensitive items)

The vault is filesystem-native — no database, no daemon. Any agent that can read Markdown can read the vault. Add your own write-gate scripts on top.

Five-minute reproducible path: see QUICKSTART.md — installs the MCP server, points it at the included sample vault, and walks through vault_search → vault_status → the readonly / full / admin safety boundary with expected output snippets.

---

MCP package

The installable MCP server lives in packages/research-vault-mcp.

npx @syndash/research-vault-mcp --transport=stdio

It exposes public-safe vault search, taxonomy, status, queue, and bounded read tools by default. Mutation tools stay hidden unless the operator explicitly widens MCP_PROFILE.

---

Live artifacts

Resource	What it is
@syndash/research-vault-mcp	Installable MCP server for Research Vault
evensong.zonicdesign.art	NAV portal — film + benchmarks + CLI showcase
evensong.zonicdesign.art/promo	2-minute promotional film (1080p)
DASH SHATTER	Agent product site
EverMind.ai	Memory & cognition research platform (related work)

---

License

Released under CC BY-NC-ND 4.0. You may share unmodified copies with attribution for non-commercial use. For commercial licensing or research collaboration, open an issue.

---

_{Provenance — The vault template originates from the Evensong research project.
Benchmark data referenced by ID (R011-B, R012-E) is for internal self-evolution reference only;
do not redistribute the underlying datasets.}

↑ Back to top