What problem does it solve?
Agents struggle to remember, retrieve, and forget information reliably over time, which causes inconsistent behavior, hallucinations, token budget waste, and failures to meet compliance or audit requirements. This Skill provides the decision logic and algorithmic tradeoffs needed to build memory systems that balance recall accuracy, latency, storage cost, and token efficiency across short- and long-term horizons.
Core Features & Use Cases
- Algorithm selection & scaling: Guidance on choosing and tuning HNSW, IVF-PQ, PQ, DiskANN, and distributed indices for different memory volumes and latency targets.
- Temporal & graph storage: Designs for bi-temporal triple stores and hybrids combining knowledge graphs with vector search for relational and multi-hop queries.
- Retrieval scoring & forgetting: Adaptive ranking combining relevance, recency, and importance plus decay-based, session-scoped, and never-forget policies and consolidation strategies.
- Failure modes & quality gates: Detection and mitigation steps for retrieval hallucination, token budget overflow, semantic drift, and knowledge graph explosion.
- Use Case: Build a personal assistant memory that serves 1M memories with sub-200ms P95 retrieval, bi-temporal auditability, and >50% storage reduction via consolidation.
Quick Start
Design an episodic memory system for a personal assistant that uses HNSW for sub-100ms retrieval, a bi-temporal knowledge graph for fact history, and decay-based forgetting policies for low-importance memories.
