Memory Layers

Documentation Index

Fetch the complete documentation index at: https://docs.memorycrystal.ai/llms.txt

Use this file to discover all available pages before exploring further.

Memory Crystal uses two shelves.

• One shelf is for what just happened.
• One shelf is for what should stay important later.

That way your AI can remember both the recent conversation and the bigger lessons.

​

What this means in practice

The two main layers are:

​

Short-term memory (STM)

This stores recent or raw conversation history. It helps with continuity, exact wording, and “what were we just talking about?” questions.

​

Long-term memory (LTM)

This stores distilled durable memories like facts, rules, decisions, goals, and workflows. It helps with cross-session continuity and high-value recall.

​

How it actually works

The repo implements these layers through:

• crystalMessages for short-term/raw message storage
• crystalMemories for durable long-term memory records

Important sources:

• README.md
• docs/00-project-overview/OVERVIEW.md
• convex/schema.ts
• convex/crystal/messages.ts
• convex/crystal/memories.ts
• convex/crystal/recall.ts

​

Commands / examples

Useful companion docs:

• Context Engine
• Memory Stores
• Knowledge Bases

​

Common mistakes

• thinking “messages” and “memories” are the same table or concept
• assuming all stored text is permanent
• forgetting that recent message continuity and durable memory retrieval solve different problems

​

Source of truth

Primary files behind this page:

• README.md
• docs/00-project-overview/OVERVIEW.md
• convex/schema.ts
• convex/crystal/messages.ts
• convex/crystal/memories.ts