Three-Layer AI Memory

Three-layer AI memory is Roberts’s structure for giving an assistant durable context: short-term memory answers “who am I?”, mid-term memory answers “what am I doing?”, and long-term memory answers “what happened before?” or “what does the expert knowledge base know?” ^[src-059].

Key points

• Short-term memory: stable identity and preference instructions such as name, role, goals, tools, tone, voice, non-negotiables, and answer style ^[src-059].
• Mid-term memory: active projects, clients, startup goals, health goals, or work domains represented by folders and project operating manuals ^[src-059].
• Long-term memory: archives of meaningful conversations plus expert knowledge bases stored in systems such as Pinecone or Obsidian ^[src-059].
• The model’s native memory is not enough for important facts because models forget, truncate, or hallucinate; important context must be written down ^[src-059].
• The pattern generalizes across Claude, Codex, Antigravity, VS Code, and other agent work surfaces when the memory lives in files or external indexes ^[src-059].

Related entities

• Jack Roberts
• Claude Code
• Codex (OpenAI)
• Pinecone
• Obsidian

Related concepts

• AI Memory Operating System
• Project Operating Manual
• Conversation Wrap-Up Memory
• Expert Knowledge Index
• Claude Code Context Management Discipline

Source references

• ^[src-059] Jack Roberts — “This Memory System just 10x’d Claude Code” (2026-05-03)