What can a memory contain?

A free-text content field plus structured metadata: category, importance (0–1), confidence (0–1), source conversation, timestamps and a vector embedding. The metadata drives scoring and the embedding drives semantic search.

How are memories scored at retrieval?

Five factors: relevance 50% (semantic + keyword match), importance 15%, confidence 15%, recency 10%, frequency 10%. Relevance dominates so the right memory beats the most recent one.

Are duplicate memories deduplicated?

Yes. Jaccard similarity ≥60% with at least 3 shared keywords triggers consolidation. Near-duplicates are merged keeping the highest confidence and most recent source.

Do memories ever expire?

Only if both unused for 120 days AND importance below 0.1. High-importance memories never expire. You can also manually delete or archive any memory at any time.

Memory System

Alma's memory system has three layers — Semantic Memories (discrete facts and preferences), Episodes (compressed conversation summaries) and Procedures (learned behavioural patterns). Every memory is indexed for hybrid retrieval (keyword + semantic) and scored by relevance, importance, confidence, recency and frequency. Retrieval happens in under 100ms.

Alma's 3-layer memory architecture: Semantic Memories capture discrete facts and preferences. Episodes store compressed conversation summaries. Procedures record learned behavioral patterns. Each memory is semantically indexed for fast retrieval via keyword, semantic, or hybrid search.

See plans