Vesper Memory

What it does

Vesper Memory gives AI agents a persistent, learning memory system that goes far beyond simple key-value storage. It connects to Claude Code (and other MCP clients) via stdio and provides a three-layer architecture inspired by human cognition: working memory for recent conversations, semantic memory with knowledge-graph-powered retrieval, and procedural memory that learns skills and workflows over time. The SmartRouter classifies queries automatically and routes them through the optimal retrieval path — whether that's a Redis cache hit, HippoRAG graph traversal, or hybrid semantic search with reciprocal rank fusion.

Tools

• store_memory — Store memories with automatic BGE-large embedding, dual storage (SQLite + Qdrant), and working memory cache. Supports namespaces for multi-agent isolation.
• retrieve_memory — Smart-routed retrieval across all memory layers. Classifies query type (factual, preference, project, temporal, skill, complex) and picks the best strategy.
• list_recent — Pull recent conversations from working memory (Redis, <5ms).
• get_stats — System health metrics: cache hit rates, entity/relationship counts, performance percentiles.
• delete_memory — Remove a memory across all layers with orphan cleanup.
• share_context — Copy memories between namespaces for multi-agent handoffs.
• store_decision — Store architectural decisions with reduced decay (4x longer retention) and conflict detection.
• list_namespaces — Discover all namespaces with memory counts.
• namespace_stats — Per-namespace breakdown of memories, entities, agents, and skills.
• load_skill — Lazy-load full skill descriptions on demand (90% token savings).
• record_skill_outcome — Track skill execution success/failure for continuous learning.
• vesper_enable / vesper_disable / vesper_status — Toggle system for A/B benchmarking.

Installation

{
  "mcpServers": {
    "vesper": {
      "command": "node",
      "args": ["~/.vesper/dist/server.js"],
      "env": {
        "REDIS_HOST": "localhost",
        "REDIS_PORT": "6379",
        "QDRANT_URL": "http://localhost:6333",
        "SQLITE_DB": "~/.vesper/data/memory.db",
        "EMBEDDING_SERVICE_URL": "http://localhost:8000"
      }
    }
  }
}

Quick install via npm:

npm install -g vesper-memory
vesper install
vesper init

Requires Docker for Redis, Qdrant, and the BGE embedding service.

Supported hosts

Claude Code (primary target, with dedicated vesper init rules installer). Other MCP clients that support stdio transport may also work.