Repository navigation
Tags
Tags
v3.0 Research-Driven Intelligence Delivered: Research-backed memory intelligence with LLM re-ranking, relation classification, spreading activation, memory fusion, and behavioral profiling — validated with 11 E2E WebSocket tests. Key accomplishments: - LLM-based search re-ranking with score blending for ~87% recall - LLM-guided relation classification with batch support and regex fallback - ACT-R/SYNAPSE spreading activation for diverse graph-based retrieval - Memory fusion merging dormant related memories into stronger summaries - Behavioral profiling with EMA-smoothed signals as natural-language insights - 11 E2E tests validating full v3.0 pipeline through WebSocket conversations 6 phases, 13 plans, 55 commits, 72 files changed. Papers: SYNAPSE, FadeMem, A-MEM + 60 more. See .planning/MILESTONES.md for full details.
v2.0 Agent Polish & Enhanced Loop Delivered: Enhanced agent with web UI, loop-until-done behavior, multi-channel triggers, and consolidated 60-line system prompt. Key accomplishments: - Removed tool system entirely (pure skills-only) - Built web UI chat interface for local testing - Implemented loop-until-done with [DONE] marker - Created TriggerSource abstraction for multi-channel dispatch - Added proactive execution guidelines - Consolidated system prompt: 217→60 lines (72% reduction) See .planning/MILESTONES.md for full details.
v1.0 Skills-Only Architecture Delivered: Complete skills-only architecture replacing hardcoded tools with discoverable skill files. Key accomplishments: - Defined SKILL.md format with YAML frontmatter and scripts folder - Implemented bash skill with command sandboxing - Built SkillExecutor class with script spawning and output capture - Created web_search, browser, memory_search, telegram_send skills - Refactored agent loop to skills-only execution See .planning/MILESTONES.md for full details.