AgentDB Memory Patterns

Initialize AgentDB with npx agentdb init, then use the createAgentDBAdapter from agentic-flow to store conversation patterns with vector embeddings. Retrieve relevant context using retrieveWithReasoning with domain filtering and MMR ranking.

Run npx agentdb mcp to start the MCP server, then register it once with claude mcp add agentdb npx agentdb mcp. This exposes AgentDB memory operations directly to Claude Code sessions.

Yes, AgentDB is 100% backward compatible with the ReasoningBank API. Use the migrateToAgentDB function or npx agentdb migrate --source command to transfer patterns with automatic validation.

AgentDB includes 9 learning plugins: Decision Transformer, Q-Learning, SARSA, Actor-Critic, Active Learning, Adversarial Training, Curriculum Learning, Federated Learning, and Multi-task Learning. Create plugins with npx agentdb create-plugin.

Large databases result from storing unquantized embeddings. Enable binary quantization for 32x memory reduction or scalar for 4x reduction, and run memory consolidation to prune low-quality patterns below a relevance threshold.

AgentDB performs vector search in under 100 microseconds using HNSW indexing and pattern retrieval in under 1ms with caching. Batch inserts of 100 patterns complete in approximately 2ms.