.agents/MEMORY.md

Current Context

Working on standardizing the Open Agent Standard and Signet AI memory system while maintaining active development on IM apps and platform integrations.

Active Projects

Signet AI / Open Agent Standard Location: Planning repo (website, landing page, docs, program sub-modules) Status: Core spec (0.2.1-draft) and architecture reviewed; key architectural decisions documented Next: Implement SDK integration with Versaal AI SDK and Anthropic AI SDK following their conventions; ensure OpenCode SDK is usable in TypeScript and Python

IMessage Viewer Location: /mnt/work/amari/imessage-viewer Status: Semantic search with nomic-embed-text pipeline operational; Archive Noir theme implemented Next: Ongoing feature development; read-only chat.db maintained with separate index.db for derived data

Compass Location: GitHub.com/High-Performance-Structures/compass Status: PR #82 (Cloudflare build fix) completed and verified Next: Continued deployment and platform compatibility work

Recent Work

Completed migration of useConversations hook from layout file to proper component structure in Compass, resolving Cloudflare build failures. Documented complete memory system architecture at spec/memory-system-design.md, including taxonomy, importance decay, daily regeneration, and platform-specific hook implementations. Integrated semantic search pipeline with nomic-embed-text (Ollama) processing 30-minute text chunks across 113K messages. Established SQLite-based memory taxonomy with FTS5 for search and local models for regeneration.

Technical Notes

• Semantic Search: nomic-embed-text via Ollama, 768-dim Float32Array BLOBs, 30-min chunk granularity
• Memory Regeneration: Local models (currently Ollama) - process is implementation detail; spec defines interface only
• SQLite Memory Taxonomy: Session (current context), Episodic (daily), Semantic (importance-decayed), Procedural (skills/recipes)
• Hooks Pattern: Platform-specific (Claude Code-equivalent, OpenClaw-equivalent) - each defines its own onSessionStart/onPrompt/onSessionEnd/onMemorySave/onMemoryQuery
• Package Manager: pacman for Arch Linux; yay for AUR packages
• Connectors: TypeScript framework, Python bindings available, follows Anthropic/Versaal SDK conventions
• Agent Identity: Portable, user-owned format; database (SQLite) as source of truth

Rules & Warnings

• Never write to or delete from source chat.db - maintain read-only, use separate index.db for derived data
• Preserve agent memory and context across sessions - user owns their agent identity
• Memory regeneration process is flexible - spec defines interface only, not implementation
• Follow platform SDK conventions when integrating Signet with Anthropic/Versaal AI SDKs
• Use importance decay in semantic memory to prioritize relevant long-term context
• Keep memory system portable and user-owned - don't lock agent identity to any single platform