NEON-SOUL - Self-Learning Soul Synthesis for AI Agents

**Changelog for neon-soul v0.4.5** - Dropped support for reading session logs (`~/.openclaw/agents/main/sessions/`). Soul synthesis now processes only memory files. - Updated documentation to reflect the new input scope, simplifying descriptions. - Cleaned up metadata: removed session log read paths from config for clarity and accuracy. - All other functionality, commands, options, and privacy guarantees remain unchanged.

- Updated version to 0.4.4. - Minor documentation and metadata changes in SKILL.md. No user-facing behavior changes documented. - No new features or options added. Existing functionality and commands remain unchanged.

- Updated version to 0.4.3. - Minor documentation and formatting adjustments in SKILL.md. - No user-facing feature or command changes.

- Documentation updated in SKILL.md with minor text and version changes for 0.4.2. - No code or behavioral changes; all updates are limited to project documentation.

- Added explicit read/write paths, minimum Node.js version, and Ollama service requirements to skill metadata. - Updated tags, paths, and options for improved clarity. - Enhanced documentation for options (e.g., clarified custom path arguments). - No changes to behavior or commands.

**Automated soul synthesis is now powered by a bundled Node.js engine with fast, incremental processing and Ollama support.** - Adds a self-contained processing script (`scripts/neon-soul.mjs`) — no more pure-instruction flows; synthesis is direct, efficient, and works offline. - Supports incremental soul synthesis by default: only new/changed memories and sessions are processed, with intelligent caching for speed and minimal resource use. - Ollama is now required locally for all synthesis; clear error guidance if not running. - Synthesis and supporting commands (`status`, `rollback`, `audit`, `trace`) are fully described and use modern, informative reporting. - The interface, commands, reporting, and documentation have been overhauled for clarity, reliability, and easy scheduling (via cron). - All privacy, provenance, and output/backup handling details updated to reflect the new architecture.

Cycle management, anti-echo-chamber protection, TOCTOU fix, Zod validation

BREAKING: Remove @xenova/transformers dependency, LLM-based semantic similarity (requires active LLM connection)

Fix 'No external code execution' contradiction: clarify local code execution vs external API calls. Add model source/integrity verification. Add Local vs External distinction table.

Fix VirusTotal flag: change homepage URL from young domain (liveneon.ai) to GitHub repo. Add embedding model fail-fast behavior with EmbeddingModelError. Add Requirements section documenting all-MiniLM-L6-v2 dependency.

Remove explicit workspace path from metadata - use relative paths only

Fix metadata format: use metadata.openclaw.config.stateDirs and requires.config for proper registry propagation

Fix metadata for security scan: add workspace path to configPaths, clarify embeddings vs LLM invocation

Add configPaths, fix disable-model-invocation registry flag

Remove skill/README.md flagged by security scan

Fix security scan warnings: add disableModelInvocation, clarify data handling, update auto-commit docs

Fix version mismatches and README paths (post-deployment review findings)

Security documentation clarifications: How This Works section, Data Access section, opt-in git integration, explicit trigger opt-in

Initial release: AI identity synthesis with semantic compression