Digital Me – Technical Whitepaper

Skills are the fundamental unit of capability in Digital Me. Each skill is a self-contained, sandboxed module that performs a specific action on behalf of the user. Skills are inspired by the Agent Skills pattern formalized by Anthropic (October 2025) and adopted as an industry standard under the Linux Foundation’s Agentic AI Foundation – but with critical security enhancements that address the supply chain vulnerabilities exposed by OpenClaw’s ClawHub disaster.

Each skill is defined by a SKILL.md manifest file containing YAML frontmatter that specifies a unique skill identifier and version, human-readable name and description, required permissions (declared capabilities), trigger patterns (semantic descriptions), input/output schemas, author identity and ShareRing verification status, and a cryptographic hash of the skill code for integrity verification.

Digital Me implements a capability-based permission system that eliminates blanket access entirely. Each skill must declare the specific ShareRing bridge methods it requires at registration time. The skill runtime enforces these declarations at execution time – a skill that declares it needs getBalance() cannot call signTransaction(), regardless of what its code attempts to do.

Digital Me ships with a comprehensive library of built-in skills. Each skill has been security audited and is maintained by ShareRing:

Identity & Verification

Financial & Wallet

Communication & Productivity

Travel & Lifestyle

Commerce & Shopping

Beyond the built-in library, Digital Me supports a curated community skills marketplace with a rigorous 6-step security review process designed to prevent supply chain attacks:

Prompt injection is the single greatest threat to AI agent security. Digital Me implements a multi-layered defense-in-depth strategy:

The most critical architectural decision in Digital Me is the separation of the data plane from the control plane at the SDK level. The ShareRing Vault SDK enforces that any operation that reads user data, shares credentials, signs transactions, or accesses encrypted storage requires explicit, PIN-gated user confirmation. This confirmation flow is implemented in the native ShareRing Me app, outside the MeModule’s WebView sandbox. Even if the AI’s context is fully compromised by a prompt injection attack, the attacker cannot read Vault documents without the user’s PIN, share credentials without the user seeing exactly what will be disclosed, sign transactions without PIN entry, or access encrypted data without PIN-gated decryption.

All data entering Digital Me from external sources (web pages, emails, messages, third-party APIs) is tagged with a taint flag. Tainted data is processed for information extraction but cannot be used as control-plane instructions. This prevents the “confused deputy” attack where malicious data sources embed instructions that the AI interprets as user commands.

Each skill executes in a sandboxed JavaScript context with access only to its declared capabilities. Skills cannot access other skills’ state, cannot modify the agent’s core memory, and cannot make network requests unless explicitly declared and approved.

Digital Me monitors for anomalous patterns that may indicate prompt injection attempts: rapid sequential skill invocations, requests for capabilities the conversation context doesn’t justify, attempts to access credentials outside the user’s normal patterns, and natural language patterns known to be associated with injection attacks.

Digital Me does not lock users into a single AI provider:

A critical design principle: the MeModule itself ships lightweight at approximately 18 MB. On-device LLM support is an optional download (1-2 GB), similar to downloading an offline language pack. For routine skill-based operations, the embedding router handles everything in under 30 milliseconds with no LLM involvement at all.

Digital Me maintains a persistent memory system using the ShareRing Me app’s AsyncStorage, enabling conversation continuity across sessions. The memory architecture is designed with privacy as the primary constraint: conversation history (last N messages, configurable, default 50) is stored locally with no server transmission; user preferences are stored as structured key-value pairs; identity cache stores non-sensitive summaries for quick access; and skill state is isolated per-skill and cleared on uninstall. All stored data is encrypted using the ShareRing Vault’s encryption capabilities.

The entire Digital Me MeModule codebase will be published on ShareRing’s GitHub repository under an open-source license upon public release. This includes the complete MeModule application code, the bridge layer implementation, the embedding router with pre-trained models, the agent core with skill orchestration logic, the complete built-in skills library, the marketplace review tooling, and comprehensive documentation with security audit reports.

While Phase 1 gives users a fully functional, privacy-first AI assistant, it leaves a fundamental tension unresolved: on-device models are limited in capability, and cloud models require trusting a centralized provider with your data. Phase 2 resolves this tension by introducing MeChain – a purpose-built Cosmos SDK blockchain that coordinates decentralized, privacy-preserving AI inference.

MeChain is a Cosmos SDK appchain built on CometBFT consensus, inheriting single-slot finality, ABCI++ VoteExtensions for validator-attested off-chain compute verification, CosmWasm smart contracts, and native IBC connectivity to 115+ chains with $58B+ market cap and zero exploits since April 2021.

MeChain extends CometBFT consensus with AI-specific validation mechanics. Validators are not merely block producers – they are compute providers who attest to the correctness of AI inference results. Target specifications: 5,000 TPS sustained throughput, sub-500ms block times, BlockSTM parallel transaction execution, and sub-second finality for inference requests.

The validation layer moves the security boundary away from the AI’s natural language understanding and into cryptographic and hardware-enforced controls. The core mechanisms include Zero Trust by Default with cryptographic confirmation at the Vault SDK level, Hardware-Level Data/Control Plane Separation, Input Sanitization and Taint Tracking, Skill Isolation and Least Privilege with capability-based permission enforcement, and continuous Anomaly Detection and Rate Limiting.

Trusted Execution Environments (TEEs) – Primary Method. TEE-based inference is the production-ready approach, offering near-native performance with hardware-enforced isolation. NVIDIA’s Blackwell B200 GPUs deliver TEE-I/O capability with inline NVLink encryption at nearly identical throughput. On MeChain, compute providers run inference within TEE enclaves. The user’s query is encrypted end-to-end; the compute provider’s operating system never sees the plaintext data.

Zero-Knowledge Machine Learning (zkML) – For Verifiable Inference. zkML enables proving that a specific model produced a specific output without revealing the inputs. zkLLM (CCS 2024) proved LLaMA-2-13B inference in approximately 15 minutes with proof size under 200 KB. Lagrange Labs’ DeepProve-1 achieved the first production-ready proof of full GPT-2 inference at 158x faster than previous approaches.

MeChain operates a decentralized marketplace where compute providers register hardware capabilities with staked SHR as collateral, a decentralized model registry provides on-chain attestation of model weights and provenance, Digital Me clients select providers based on latency, cost, TEE certification, and stake amount, validators periodically challenge providers to verify inference quality, and IBC interoperability connects to Akash Network, Nillion, and ShareLedger.

The Digital Me ecosystem introduces a dual-token model that separates the identity and trust layer from the compute layer, with each token serving a distinct and complementary function.

The dual-token model is designed so MeToken increases the utility and demand for SHR through reinforcing mechanisms: every MeChain operation requires a verified ShareLedger identity (DID) – more MeChain activity drives more ShareLedger transactions; all skills marketplace earnings flow through SHR; cross-chain verification fees consume SHR; and SHR holders govern the foundational identity layer that MeChain depends on.

To bootstrap the MeChain ecosystem, ShareRing will conduct a one-time airdrop of MeToken to participants actively holding or staking SHR on ShareLedger at snapshot time. Eligible: SHR holders and stakers with tokens on ShareLedger (native chain), including wallet holdings, validator delegations, and governance participation. Not eligible: SHR held on centralized exchanges or as ERC-20 tokens on Ethereum. Only native ShareLedger holdings qualify.

A dedicated bridge contract enables immediate, frictionless conversion between SHR and MeToken via IBC transfer, with conversion rates determined by an on-chain automated market maker (AMM) pool ensuring continuous liquidity and transparent pricing.

Phase 3 transforms Digital Me from a reactive assistant into a proactive agent that learns from and adapts to each user through on-device RAG and federated LoRA fine-tuning – all while maintaining strict privacy guarantees.

Digital Me constructs a personal knowledge graph from user interactions, capturing entities, relationships, temporal patterns, and preferences – stored entirely on-device in an encrypted SQLite database with vector indexing, never transmitted to any server. The Adaptive RAG Pipeline implements Google AI Edge’s on-device framework with automatic privacy-aware chunking, GTE-small embedding, HNSW-indexed vector search for sub-millisecond retrieval, and intelligent context window management (32K tokens for Gemma 3n).

LoRA enables personalizing an LLM by training small adapter matrices (1-10 MB) rather than modifying full model weights. Digital Me continuously trains a personal LoRA adapter on-device during idle periods, learning communication style, frequently requested information, and typical workflows. The adapter is stored in the user’s encrypted Vault – representing the user’s “personality” for the AI.

Users who opt in to federated learning contribute anonymized adapter updates to the MeChain network using state-of-the-art techniques including FlexLoRA for heterogeneous-rank contributions, FedBuff for 3.3x more efficient async aggregation, ECOLORA for 40-70% bandwidth savings, differential privacy at ε < 10 with strong downstream performance, and Ditto Regularization for jointly optimized global and personalized models.

With RAG and LoRA personalization, Digital Me transitions to a proactive agent: personalized morning briefings, smart notification filtering, automated workflow suggestions, contextual credential preparation at known venues, financial spending insights, document lifecycle management, and predictive travel assistance – all assembled on-device.

Phase 4 represents the full realization of Digital Me: an autonomous agent that can negotiate, transact, and represent you in the digital economy.

Users issue W3C Verifiable Credentials to their Digital Me agent specifying scoped permissions, anchored on ShareLedger. Each delegation credential specifies the scope of permitted actions, financial limits, time-bounded validity, revocation conditions, and required user confirmation thresholds. Service providers verify the delegation chain back to the user’s DID.

Digital Me agents communicate using Google’s A2A protocol for discovery and task negotiation, Anthropic’s MCP for tool and data access, and DIDComm v2.1 for secure messaging. Scenarios include automated hotel booking with credential verification, credential-gated access to professional services, decentralized marketplace transactions with CosmWasm escrow, and cross-agent collaboration for group planning.

MeChain’s IBC connectivity enables Digital Me to operate across the entire Cosmos ecosystem and, via IBC v2 bridges, into Ethereum and Solana ecosystems – presenting verified identity on DeFi protocols, using credentials on freelancer marketplaces, or proving nationality on travel platforms from a single, user-controlled identity.

Phase 4 enables a new economic model where users are active participants: data sovereignty with monetizable anonymized insights, skill authorship earnings in SHR, compute provision on MeChain, federated learning rewards, and configurable agent services (e.g., a professional translator’s Digital Me offering verified translation).

Digital Me’s security architecture is designed from first principles to prevent the classes of attacks that devastated OpenClaw. Every design decision assumes the AI’s context will be compromised and ensures that compromise cannot lead to unauthorized data access or transactions.

The ShareRing Vault SDK provides the hardware-backed security boundary that OpenClaw lacks entirely. All sensitive operations require explicit user PIN confirmation through the native app, outside the MeModule’s WebView. There is no API to suppress the confirmation dialog, auto-approve operations, or bypass PIN verification – these are hardcoded architectural constraints, not configuration options.

Digital Me implements automated red teaming with multi-turn prompt injection probes, dependency scanning with immediate patch deployment, a public bug bounty program covering all components, regular third-party security audits with published results, and on-device anomaly monitoring for unusual patterns.

Digital Me’s unique competitive advantage is the combination of three capabilities no competitor possesses: (1) production self-sovereign identity infrastructure with regulatory certification (UK DIATF, Australian government age assurance trial), (2) privacy-preserving AI with layered techniques from on-device to TEE to ZK, and (3) an extensible, open-source skills marketplace with security guarantees. The EU AI Act reaches full applicability in August 2026, and Digital Me’s privacy-by-design architecture positions it as compliant by default.

Digital Me is designed for compliance across the world’s most stringent jurisdictions:

Subject to change