About this skill

Problem

AI agents often fail not because they miss one fact, but because memory is fragmented, retrieval is imprecise, sync is inconsistent, and the system cannot judge which experiences deserve long-term recall. A plain vector store answers “what is similar,” but not “when to surface this memory” or “which lesson is worth keeping.”

How It Works

memory-orchestrator organizes memory into indexing, sync, understanding, structuring, and feedback loops:
- Semantic search: uses FAISS, all-MiniLM-L6-v2, and qwen2.5:7b for natural-language retrieval with fuzzy matching and context awareness.
- Multimodal indexing: converts images and audio through CLIP and Whisper, then embeds them into the same search space.
- Sync and encryption: combines Syncthing, Git, and git-crypt for offline-first cross-device sync with encrypted sensitive files.
- Structuring and recall: builds entity-relation graphs with NetworkX and PyVis, while a trigger engine surfaces relevant memories by keyword, time, or scenario.
- Self-optimization: applies scoring and A/B testing to tune retrieval, archive low-value entries, and improve future recall.

Boundaries

It is best suited to local personal knowledge bases, project retrospectives, writing asset management, and emotion journals. It still depends on a correctly deployed Ollama setup, models, Syncthing, Git, and related tools; emotion labels and graph quality depend on extraction and classification heuristics, so high-stakes decisions should not rely on automatic recommendations alone.

Screenshots

Use Cases

Engineers who need fuzzy natural-language retrieval over scattered notes, decisions, and postmortems
Independent developers maintaining an encrypted, offline-first knowledge base across multiple devices
Creators who want image and audio assets indexed automatically and linked to related past ideas
Tech leads extracting project lessons and entity relationships for retrospectives and future recall

Best For

Local AI agent developers who need retrieval, sync, graphs, and self-optimization in one system
Personal knowledge base maintainers who need encrypted cross-device sync and natural-language recall
Creators and researchers who need image and audio assets turned into searchable, recommendable indexes
Retrospective-driven engineers who need to extract lessons and improve future retrieval strategies