reasoningbank-agentdb_skill

Overview

This skill implements ReasoningBank adaptive learning on top of AgentDB's high-performance vector database. It combines trajectory tracking, verdict judgment, memory distillation, and pattern recognition so agents can learn from past experiences and improve decisions over time. The integration preserves legacy ReasoningBank APIs while delivering sub-millisecond retrieval and large-speedup batch operations.

How this skill works

The adapter stores experience and trajectory embeddings in AgentDB, enabling rapid similarity search and reasoning. Retrieval APIs return ranked memories and synthesized context, and optional reasoning modules (PatternMatcher, ContextSynthesizer, MemoryOptimizer, ExperienceCurator) produce verdicts, distilled patterns, or consolidated memories. Memory operations support automatic optimization, pruning, and hierarchical abstraction for transfer learning across domains.

When to use it

• Building self-learning agents that must replay and learn from past executions.
• Implementing experience replay or reinforcement-learning pipelines with fast memory access.
• Automating verdicts on action sequences or judging trajectory success.
• Consolidating repeated experiences into high-level patterns for faster planning.
• Scaling memory storage and retrieval where sub-millisecond latency matters.

Best practices

• Embed trajectories, experiences, and distilled patterns consistently (same dimension) to ensure reliable similarity.
• Enable context synthesis (ContextSynthesizer) when you need richer, coherent explanations rather than raw matches.
• Use MemoryOptimizer periodically or on insert with optimizeMemory to prune low-quality records and consolidate duplicates.
• Set domain and confidence filters to reduce noise and focus retrieval on relevant, high-quality memories.
• Keep a moderate cacheSize for hot patterns and rely on AgentDB for large-scale cold storage.

Example use cases

• Record API optimization steps as trajectories, then judge new optimization plans against successful past runs.
• Distill repeated debugging fixes into mid-level patterns to accelerate future triage and fixes.
• Use experience replay in RL agents with fast batch inserts and retrieval for training iterations.
• Automatically prune and consolidate telemetry-driven operational incidents into actionable principles.
• Transfer backend optimization patterns into frontend contexts by adapting distilled memos across domains.