Home

hermes-cashew overview

hermes-cashew is a memory provider plugin for Hermes Agent that stores conversation context in a local Cashew thought graph with semantic search and automatic context recall. It bridges the Hermes Agent runtime with Cashew's SQLite-backed, locally-embedded knowledge graph.

The plugin provides persistent thought-graph memory with local embeddings, organic decay, and autonomous think cycles. It exposes two LLM-accessible tools (cashew_query and cashew_extract), runs background graph consolidation via a cron-scheduled sleep cycle, and optionally uses an LLM for structured knowledge extraction, cross-domain insight generation, and conversation-arc pattern detection.

Key capabilities:

• Three-tier retrieval: sqlite-vec semantic search with graph BFS traversal and keyword fallback
• LLM-powered extraction: structured knowledge extraction with typed nodes, confidence scores, and domain assignment
• Automatic background sync: conversation turns are persisted asynchronously via a bounded queue and non-daemon worker thread
• Sleep cycle: cron-scheduled graph consolidation pipeline with cross-linking, dedup, garbage collection, and core memory promotion
• Think cycles: periodic cross-domain insight generation
• Pre-compress insight extraction: conversation-arc pattern detection before context compression
• Queue prefetch: anticipatory memory warmup for the next turn
• Priority-based retrieval: context recall with flexible filtering, domain separation, and privacy controls

Quick links

• Architecture
• Getting started
• Glossary

License

Apache 2.0 -- see LICENSE.

---

Generated by Factory