Why three layers and not one?

Facts, events and workflows answer different questions. Mixing them in one bucket means worse retrieval. Separating them lets the AI pull the right shape of context — preferences for "how do I like X?", episodes for "what did we decide last week?", procedures for "how do we do X here?".

How is context assembled?

Parallel fetch from all three layers with a 3-second per-layer timeout, then scoring by five factors (relevance, importance, confidence, recency, frequency) and finally injection into the system prompt — all in under 100ms.

What are the scoring weights?

Relevance 50%, importance 15%, confidence 15%, recency 10%, frequency 10%. Relevance dominates so the right memory beats the most recent one.

Can I see what is in each layer?

Yes. Each layer has its own UI in Alma — memories, episodes and procedures are independently browsable, editable and exportable.

Alma's Three-Layer Memory Architecture

2026-01-28 · Alma Team

Alma stores context in three layers because not all knowledge is the same shape. Memories hold facts and preferences. Episodes capture event sequences and timelines. Procedures encode workflows and decision trees. Context assembly hits all three in parallel and builds the system prompt in under 100ms.

Not all context is equal. Memories store facts and preferences. Episodes capture event sequences and timelines. Procedures encode workflows and decision trees. Context assembly fetches all three in parallel with 3-second timeouts, scores by 5 factors, and builds the system prompt in under 100ms.

Read the full article · Alma vs ChatGPT Memory · See plans

See plans