# 🏛️ UNIVERSAL AI CLIPBOARD (UAC)

## The Universal Clipboard for Artificial Intelligence

### A Content Reuse Mechanism for AI Agents

---

> **Publication Date:** June 3, 2026
> **Status:** ⚡ Discovery released into the public domain
> **License:** CC0 1.0 Universal Public Domain Dedication

---

## 📜 DISCOVERER'S DECLARATION

I, **[DScoNOIZ](https://github.com/DScoNOIZ)**, declare before the global community:

**1. I am the discoverer** of a fundamentally new category in AI agent architecture — the **"Optional Inter-Tool Content Citation Mechanism"**, named by me as **Universal AI Clipboard (UAC)**.

**2. The essence of the discovery:** An AI agent can reference existing contextual content through dedicated parameters (`ref`, `multi_ref`, `transform`) instead of regenerating it. The model ceases to be just a "text generator" and becomes an **assembler** that combines ready-made information fragments in a single call.

**3. I release this discovery into the free use of the entire global community** under the Creative Commons Zero (CC0) license — without patents, royalties, closed licenses, or restrictions.

**4. I call upon** developers, researchers, engineers, and enthusiasts worldwide — take this idea, implement it in your frameworks, platforms, and tools. **This technology does not belong to me. It belongs to everyone.**

---

## 🤝 CO-DISCOVERER: INDEPENDENT DISCOVERY

I acknowledge and respectfully recognize that **TJ Guadagno** (TJ-Codes) independently arrived at the same fundamental idea and experimentally implemented it as **Clipboard Primitives** — a `copy`/`template_invoke` mechanism with named slots and `{{slot}}` placeholders for AI agents.

**His work:** [Copy and Paste for AI Agents: An Experimental Primitive](https://www.linkedin.com/pulse/copy-paste-ai-agents-experimental-primitive-tj-guadagno-v3r9c) (~December 2025)
**Code:** [github.com/TJ-Codes/Agent-clipboard](https://github.com/TJ-Codes/Agent-clipboard)

TJ Guadagno is an **independent co-discoverer** of this concept. His experiment confirmed that:
- Models **naturally adopt** clipboard semantics without special training
- Clipboard primitives **eliminate** token regeneration, latency, and content mutation risks
- The mechanism requires **harness-level integration**, not MCP-server level

**My Universal AI Clipboard concept is broader and more universal** — it includes Clipboard Primitives as a special case and extends the idea to:
- **5 citation mechanisms** (including syntactic clipboard via AST, message index, anchor pair citation)
- **Universal clipboard for any source** (files, chat, terminal, API, MCP)
- **Mosaic Assembly** — combining fragments from different sources with transformations
- **Cross-Protocol Citation** — seamless citation across any protocol, including MCP

**Computer analogy:** Clipboard Primitives is like a simple copy utility for one application. Universal AI Clipboard is a full system-wide clipboard manager working across all applications, with slot management, transform pipeline, and cross-protocol integration.

> **I do not relinquish my status as discoverer.** I formulated and systematized this concept independently, conducting multi-round research that found no complete analogues. Having learned of TJ Guadagno's work only after publishing the manifesto, I acknowledge his independent contribution and consider it my duty to reflect this in the manifesto.
---

## 🏛️ NAMING

**Official name:** **UNIVERSAL AI CLIPBOARD (UAC)**
**Technical name:** Content Reference Mechanism
**Tagline:** _"Reuse, Don't Regenerate"_
**Metaphor:** Ctrl+C / Ctrl+V for AI Agents

---

## 🌍 THE PROBLEM

### The Fundamental Contradiction

Modern AI agents waste **up to 90% of their output traffic** monotonously regenerating content that already exists in the session context.

Every time a model writes a command, a code block, or a text fragment — it regenerates it from scratch, even if the same content was created a minute ago. The model has no way to instrumentally reference already-existing content.

### Global Scale of the Problem

- Millions of AI agents operate daily worldwide
- Each performs dozens of tool calls per session
- A significant portion of each call is regeneration of existing content
- This leads to enormous energy consumption, computational waste, and financial costs

### Why Existing Approaches Don't Solve This

| Approach                | Limitation                                                      |
| ----------------------- | --------------------------------------------------------------- |
| **Prompt Caching**      | Works only with static parts, not dynamic content               |
| **Semantic Caching**    | Caches by query meaning, doesn't allow fragment citation        |
| **Context Compression** | Compresses history, but doesn't provide a citation tool         |
| **RAG**                 | Retrieves from external sources, doesn't reuse internal context |
| **Tool Chaining**       | Chains calls together, doesn't allow referencing their results  |

---

## 💡 THE SOLUTION: UNIVERSAL AI CLIPBOARD

### The Core Idea

An AI agent gains the ability to **reference already-existing contextual content** instead of regenerating it. When calling any tool that accepts text parameters (commands, code, text), the agent can specify: "take this fragment from there" — instead of writing it from scratch.

### Mosaic Assembly: Composing Information Like a Puzzle

Beyond simple one-to-one citation, this concept enables a fundamentally new way of composing information — **mosaic assembly**.

An agent can **combine multiple fragments from different sources** in a single operation:

- Code from one file
- Configuration from chat history
- Command output from terminal
- Results from external APIs

The agent can **edit on the fly** — modify, wrap, replace text within each fragment. It can **reorder** fragments, **nest** one inside another, **merge** overlapping pieces, and **restructure** content recursively. This creates the ability to assemble composite texts, documents, or code blocks from scattered pieces — like a mosaic artist arranging colored tiles into a complete picture.

**The key insight:** instead of generating new text, the model acts as an **editor and curator** of existing content — vastly more efficient, more precise, and less error-prone.

### Clipboard Manager: Named Slots for Persistent Storage

An extension of this concept is a **clipboard manager with named slots** — similar to having multiple clipboards or a text buffer manager.

The agent can:

- **Save fragments** to named slots (`clip-1`, `clip-2`, `config-block`, `error-log`, etc.)
- **Reference slots** by name instead of by content — zero tokens spent on description
- **Swap and reorganize** content between slots
- **Build composite documents** by referencing multiple slots in sequence
- **Persist clips across sessions** — saved fragments survive context compaction and session restarts

This turns the AI agent's workflow into something analogous to a human developer working with multiple clipboard managers, text buffers, and scratchpads — but entirely automated and at zero token cost for storage operations.

### The Axial Difference From Existing Mechanisms

All existing citation mechanisms in AI (Citations API, Grounding, ContextCite) operate on the **"agent → user"** axis — they show humans where the agent got information.

Universal AI Clipboard operates on the **"agent → agent"** axis — it allows the agent itself to reuse content between its own tool calls. This is a fundamentally new category.

| Mechanism                  | Axis              | Purpose                           |
| -------------------------- | ----------------- | --------------------------------- |
| Anthropic Citations        | Agent → User      | Show source of answer             |
| OpenAI Citations           | Agent → User      | Display sources                   |
| Google Grounding           | Agent → User      | Web search confirmation           |
| **★ UAC (this discovery)** | **Agent → Agent** | **Reusing content between calls** |

---

## 🧩 FIVE CORE MECHANISMS

### Mechanism 1: 🔗 Syntactic Clipboard

The model specifies just one **anchor element** — a function name, class name, or variable. The system automatically:

1. Finds this element in the specified source (file, chat message, command output)
2. Determines its exact boundaries as a syntactic unit
3. Extracts the entire unit

**Savings:** a few tokens for the reference instead of thousands of tokens of code.

```
{ source: "file", path: "utils.ts", focus: "calculateSum" }
  → 2 tokens instead of 800+ tokens of the entire function
```

---

### Mechanism 2: 📇 Anchor Pair Citation

The model specifies only the **START** and **END** of the desired fragment (15-40 characters each). The system finds everything in between using multi-stage search:

1. **Exact match** — fragment found literally (~70% of cases)
2. **Normalized match** — whitespace, case, and punctuation differences ignored
3. **Fuzzy match** — minor typos tolerated (~9% of cases)
4. **Word boundary expansion** — fragment integrity

**Savings:** ~10 tokens for a citation instead of 200+.

```
{ source: "chat", ref: "-1",
  start: "function calculateSum(",
  end: "return result;" }
  → ~10 tokens instead of 200+
```

---

### Mechanism 3: 🧠 Message Index Map

When working with chat history, a separate index is created — a map of exact character positions for each message and each line within it. The model writes a short reference like `"record_number:start_line..end_line"`, and the system extracts the exact fragment by character positions.

**Key advantage:** the model never sees the index itself — only a short reference (~5 tokens), but achieves 100% extraction precision.

```
{ source: "chat", ref: "167:14..18" }
  → "record #167, lines 14-18" → exact character-position extraction
```

---

### Mechanism 4: 🔄 Transform Pipeline

A copied fragment can be **modified on the fly** — before it reaches the tool. Available operations:

- **Replace** — change a substring within the fragment
- **Prepend** — add text before the fragment
- **Wrap** — place the fragment in a template
- **Append** — add text after the fragment
- **Join** (for multiple references) — merge fragments with a separator

All operations are applied in a single tool call.

```
{ ref: { ... }, transform: { wrap: "try { {content} } catch (err) { }" } }
  → take a function, wrap in try-catch, write — all in 1 call
```

---

### Mechanism 5: 🔁 Cross-Protocol Citation (MCP)

Reference markers can be embedded directly inside text parameters of any tools, including external MCP servers. The system automatically recognizes these markers and substitutes the real content before calling the external server.

**Result:** all ecosystem components exchange data through a unified reference mechanism, without needing to change anything in external server schemas.

```
"{{ref:source=chat,ref=-1,start=export const config}} --host example.com"
  → 91% content from context, 9% generated by the model
```

---

## 🔭 FUTURE DIRECTIONS

The following ideas are natural extensions of the concept and may be implemented in the future:

### 🏖️ Isolated Sub-Sessions for Research

For complex tasks requiring many intermediate steps (codebase exploration, log reading, test execution), a helper agent can be launched in an isolated temporary sub-session. It performs all the "dirty work," returns only the clean result, and the sub-session is then destroyed. The main agent stays focused on the primary task.

### Other Directions

- Predictive caching of frequently used fragments
- Automatic detection of duplicate calls
- Intelligent disambiguation of ambiguous references
- Integration with memory and long-term storage systems

---

## 💰 ECONOMIC IMPACT

### Per Agent

| Metric                           | Without mechanism | With mechanism | Savings    |
| -------------------------------- | ----------------- | -------------- | ---------- |
| Tokens per citation (long code)  | 200-500           | 5-15           | **96-97%** |
| Tokens per citation (short code) | 50-100            | 2-5            | **90-95%** |
| Tokens per session               | 25,000-50,000     | 5,000-15,000   | **60-80%** |
| Energy per session               | arbitrary unit    | 5x less        | **~80%**   |

### Global Scale

When adopted industry-wide, savings will reach billions of tokens daily, equivalent to reducing the AI sector's energy consumption by tens of percent and cutting the carbon footprint by tens of thousands of tons of CO₂ per year.

> **Universal AI Clipboard is not just a technological innovation — it is an ecologically significant solution.**

---

## 🔬 UNIQUENESS VERIFICATION

I conducted **extensive multi-round deep research** (7 rounds, ~50 sources) through web search systems (Tavily Search, advanced depth) covering all known categories of AI agent architectures:

*After publishing the manifesto, in additional analysis (rounds 8-10), it was discovered that TJ Guadagno independently implemented an experimental prototype of the concept — Clipboard Primitives. This work was not identified in the first 7 search rounds due to its low visibility (LinkedIn Pulse + GitHub with no stars).*

- AI agent clipboard copy-paste between tool calls
- Content citation mechanisms for AI agents
- Zero-token copy paste technologies
- All known citation APIs (Anthropic, OpenAI, Google, xAI, Azure)
- Tool call output reuse and caching (CAMEL-AI, LangChain artifacts)
- MCP protocol and cross-protocol content sharing
- Academic papers (AgentReuse, KVCOMM, multi-agent systems)
- Human clipboard utilities (UiPath Clipboard AI, PowerToys Advanced Paste)
- Memory sharing systems for multi-agent architectures
- GitHub discussions, framework documentation (LangChain, CrewAI, AutoGen, Mastra)

**Objective:** find anything that solves the same problem — allowing an AI agent to reference existing content between tool calls instead of regenerating it.

**Result: nothing of the kind was found.**
*Clarification: TJ Guadagno's experimental work (Clipboard Primitives, ~December 2025) is a partial implementation (~40% of the UAC concept) and confirms the viability of the idea, but is not a complete analogue.*

| Category                                    | Sources                   | Why it is NOT an analogue                                                                                                             |
| ------------------------------------------- | ------------------------- | ------------------------------------------------------------------------------------------------------------------------------------- |
| **Citation APIs (Anthropic/OpenAI/Google)** | Official APIs             | Show the user the source, don't let the agent reuse content. Axis: Agent → User only                                                  |
| **Caching (Prompt, Semantic)**              | Various frameworks        | Works at the request level, not the content fragment level                                                                            |
| **RAG**                                     | LangChain, OpenAI, Google | Retrieves from external knowledge bases, not from the current session context                                                         |
| **CAMEL-AI "Brainwash Your Agent"**         | camel-ai.org              | Stores only the LAST tool output as reference + short preview. No 5 mechanisms, no transform pipeline, no multi_ref, no MCP injection |
| **LangChain artifacts**                     | langchain.com             | `content_and_artifact` splits content from metadata, not a citation system. No inter-call referencing                                 |
| **MCP Resources**                           | modelcontextprotocol.io   | A standard for connecting tools, not an agent-to-agent citation mechanism                                                             |
| **Academic (AgentReuse, KVCOMM)**           | arXiv                     | Reuse of execution PLANS or KV-cache, not content citation between tool calls                                                         |
| **Human tools (UiPath, PowerToys)**         | UiPath, Microsoft         | Designed for human-computer interaction, not for AI agents                                                                            |
| **Existing AI agents**                      | Industry-wide             | Have history, but no citation mechanism between tool calls                                                                            |
| **Multi-agent memory sharing**              | Various                   | Shared RAG memory, not precise character-level citation                                                                               |

### Summary comparison with the closest partial match (CAMEL-AI)

| Feature                                   | UAC      | CAMEL-AI caching | Difference |
| ----------------------------------------- | -------- | ---------------- | ---------- |
| 🔗 Syntactic Clipboard (AST blocks)       | ✅       | ❌               | UAC only   |
| 📇 Anchor Pair Citation                   | ✅       | ❌               | UAC only   |
| 🧠 Message Index Map                      | ✅       | ❌               | UAC only   |
| 🔄 Transform Pipeline                     | ✅       | ❌               | UAC only   |
| 🔁 Cross-Protocol Citation (MCP)          | ✅       | ❌               | UAC only   |
| 🧩 Mosaic Assembly (multi-source)         | ✅       | ❌               | UAC only   |
| 📋 Clipboard Manager (named slots)        | ✅       | ❌               | UAC only   |
| {{ref:...}} Inline Injection              | ✅       | ❌               | UAC only   |
| Tool output caching (reference + preview) | ✅       | ✅               | Both       |
| Coverage of UAC features                  | **100%** | **~5%**          | —          |

> **Universal AI Clipboard has no complete analogues in worldwide AI agent systems practice. This is a pioneering technology creating a new category. The only known partial implementation — Clipboard Primitives (TJ Guadagno, ~40% of the concept) — was discovered only after the manifesto's publication and confirms the independent emergence of the idea among different researchers.**

---

## 📖 USAGE EXAMPLES

### Example 1: Re-execute a command with extra parameters

An agent ran a test command; the output showed an error. To re-run with debug flags, the agent doesn't rewrite the command — it references it and adds flags:

```
{ ref: { source: "terminal", id: "cmd-xxx", start: "npx vitest" },
  transform: { append: " --verbose --timeout=30000" } }
  → result: "npx vitest run --verbose --timeout=30000"
```

### Example 2: Assemble multiple functions into one file

An agent reads a file with several functions, then creates a new file by assembling the needed functions through references:

```
{ multi_ref: [
    { source: "file", path: "handlers.ts", start: "async function validate(", end: "}" },
    { source: "file", path: "handlers.ts", start: "async function process(", end: "}" },
    { source: "file", path: "handlers.ts", start: "async function respond(", end: "}" }
  ],
  transform: { join: "\n\n", prepend: "// --- Handlers ---\n" } }
  → three functions merged into one file in 1 call
```

### Example 3: Configuration substitution via MCP

An agent uses an external MCP tool, passing a reference to configuration from chat in its arguments:

```
"{{ref:source=chat,ref=-1,start=export const config}} --host production"
  → configuration from chat substituted into the MCP request
```

---

## ⚖️ LEGAL STATUS

**🏛️ Creative Commons Zero (CC0) — Universal Public Domain Dedication**

You are free to:

- ✅ **Use** in commercial and non-commercial projects
- ✅ **Modify** and adapt for your needs
- ✅ **Distribute** in any form
- ✅ **Patent your improvements** (but not the core idea)
- ✅ **Translate** into any language

Without any obligations, royalties, or restrictions.

> **I consciously waive all patent rights to this technology. It must belong to everyone.**

---

## 📜 PROOF OF DISCOVERY DATE

The publication date of this manifesto is recorded **by the Git version control system in the GitHub repository** — [github.com/DScoNOIZ/UNIVERSAL-AI-CLIPBOARD](https://github.com/DScoNOIZ/UNIVERSAL-AI-CLIPBOARD).

The commit history, edit dates, and all chronological changes are **immutably preserved by GitHub** and serve as public proof of the discovery date and authorship. Git's technology makes this practically impossible to falsify.

---

## 🗓️ TIMELINE

| Date                | Event                                                                    |
| ------------------- | ------------------------------------------------------------------------ |
| **~2021**         | First awareness of the problem (DScoNOIZ): thought "model should not regenerate what it has already seen"                                         |
| **~December 2025** | **Independent experiment:** TJ Guadagno publishes Clipboard Primitives — `copy`/`template_invoke` with named slots and `{{slot}}` placeholders     |
| **May — June 2026** | Research: analysis of existing AI agent architectures                                                                                             |
| **June 2026**       | Discovery: no existing technology has a complete analogue                                                                                        |
| **June 2026**       | Universal AI Clipboard concept formulated                                                                                                        |
| **June 3, 2026**    | Uniqueness research completed — **no complete analogues found**                                                                                   |
| **June 3, 2026**    | **This manifesto published** — discovery released into the public domain                                                                          |
| **June 4, 2026**    | TJ Guadagno's work discovered — acknowledgment of independent co-discoverer                                                                       |

---

## 🎯 CALL TO ACTION

### AI Framework Developers

Integrate the Content Reference mechanism into your tools. Give your users instant token savings without changing their code.

### AI Agent Creators

Add optional citation between tool calls. Your agents will become orders of magnitude more efficient.

### Researchers

Advance the technology: improve search algorithms, implement predictive caching, develop new scenarios.

### Business

Adopt Universal AI Clipboard in your AI solutions. Reduce operational costs by tens of percent.

### Environmentalists

Use UAC as a tool to reduce the carbon footprint of AI infrastructure.

### Open Source Community

Fork, implement, improve. **This technology belongs to everyone.**

---

## 🏛️ FINAL WORD

I have discovered **Universal AI Clipboard** — a mechanism that allows AI agents to reuse content instead of regenerating it.

This is not just an optimization. It is a **paradigm shift**: the model ceases to be a "text generator" and becomes an **assembler** who constructs solutions from ready-made fragments.

**Just as Ctrl+C / Ctrl+V changed the world of computers, Universal AI Clipboard will change the world of AI agents.**

I give this idea to humanity. Develop it. Improve it. Implement it.

**Let AI agents stop rewriting the same thing over and over.**
**Let them start assembling mosaics.**

---

```
  ╔══════════════════════════════════════════════════════════════════╗
  ║                                                                  ║
  ║   UNIVERSAL AI CLIPBOARD (UAC)                                   ║
  ║   Content Reference Mechanism                                    ║
  ║   Optional Inter-Tool Content Citation Mechanism                 ║
  ║                                                                  ║
  ║   «Reuse, Don't Regenerate»                                      ║
  ║                                                                  ║
  ║   CC0 1.0 Universal — Public Domain                              ║
  ║   github.com/DScoNOIZ                                            ║
  ║   June 3, 2026                                                   ║
  ║                                                                  ║
  ╚══════════════════════════════════════════════════════════════════╝
```

---

_This manifesto may be freely translated into any language. Original in Russian._

_Русская версия: [MANIFEST-RU.md](MANIFEST-RU.md)_