hierarchical-memory

Use a layered memory architecture: store session context in a short-term in-memory store, persist important patterns in a Qdrant vector database, and log decisions as episodic records. Consolidation hooks migrate key information at session end.

Short-term memory holds current session context with token-based search and a 100-item limit. Long-term memory stores permanent knowledge in Qdrant with vector similarity search. Episodic memory records timestamped decisions and events in JSONL files retained for 30 days.

Text is converted to vector embeddings and stored in a Qdrant collection on localhost:6333. Queries are embedded the same way, and Qdrant returns the most similar stored entries, enabling retrieval even when wording differs from the original.

First verify Qdrant is running by checking localhost:6333/health, then confirm the target collection exists. If retrieval is still empty, make the search query more specific or check whether consolidation hooks ran at session end.

Vector search can return loosely related results for vague queries, so metadata tags like topic and priority should be added when storing. Outdated patterns also accumulate over time and require periodic cleanup to keep retrieval accurate.