Mnemosyne

Mnemosyne

High-performance agentic memory system for Claude Code's multi-agent orchestration

Mnemosyne provides persistent semantic memory with sub-millisecond retrieval, built in Rust with LibSQL vector search and PyO3 Python bindings.

Features

Core Memory System
- Project-Aware: Automatic namespace detection from git repositories and CLAUDE.md
- Semantic Search: LibSQL vector embeddings + full-text search (FTS5) + graph connectivity
- Type System: Insight, Architecture, Decision, Task, Reference memory types
- Graph Linking: Automatic bidirectional relationship management
- Privacy-First: Local-only storage with optional privacy-preserving evaluation

Multi-Agent Orchestration
- Ractor Actors: 4 specialized agents (Orchestrator, Optimizer, Reviewer, Executor)
- LLM-Enhanced Reviewer: Automatic requirement extraction, semantic validation, intent verification with Claude API
- Work Queue: Dependency-aware scheduling with priority management
- Quality Gates: Automated test verification, anti-pattern detection, constraint validation, requirement traceability
- Deadlock Resolution: Priority-based preemption (60s timeout)
- Sub-Agent Spawning: Parallel work execution across child actors
- Event Persistence: Complete audit trail of orchestration events with SSE broadcasting

Distributed Coordination
- Peer Discovery: Automatic peer discovery on local network via mnemosyne peer invite/join
- Work Delegation: Seamless offloading of tasks to available peers
- Iroh Networking: P2P encrypted communication layer for secure direct connections

Network Visualization
- Network Graph: Visual representation of connected peers and topology via mnemosyne graph
- Real-time Status: Connection latency, bandwidth, and peer health monitoring
- Topology Awareness: Automatic detection of…