Memory

Platforms: claude openai gemini m365-copilot

What Memory Is

Memory is accumulated knowledge from past interactions — preferences, decisions, facts, and patterns that the AI retains and retrieves when relevant. Memory makes AI persistent rather than stateless: instead of starting from scratch every conversation, the AI improves over time.

The key distinction from Context: Context is knowledge you provide; Memory is knowledge the system accumulates. You curate context by uploading files, pasting examples, or configuring project knowledge. Memory is managed by the AI itself — it decides what to remember based on your interactions.

Key Characteristics

• System-managed, not user-curated — the AI decides what to remember from your interactions
• Persists across conversations — survives session boundaries, available in future interactions
• Grows over time — more interactions produce richer memory, improving personalization and relevance
• Two main types — short-term (within a session) and long-term (across sessions)

When to Use Memory

Use memory when:

• Repeating context to the AI is friction — preferences, project conventions, communication style
• Personalization improves the experience — the AI should adapt to how you work
• The AI needs to learn from past interactions — building on previous decisions and patterns
• Continuity matters — picking up where you left off without re-explaining

When memory isn’t available or appropriate, use Context to provide knowledge explicitly for a single conversation or project.

Platform Implementations

Platform	How It Works
Claude	Claude memory, CLAUDE.md project memory, conversation continuity
OpenAI (ChatGPT)	ChatGPT Memory, custom instructions persistence
Gemini	Conversation memory, Gems with learned preferences
M365 Copilot	Microsoft Graph as implicit memory, organizational knowledge

Types of Memory

Short-term memory (working context)

The current conversation history and documents loaded into the context window. This is what the AI can “see” right now — it’s limited by the model’s context window size and disappears when the conversation ends.

Long-term memory (persistent storage)

Information stored outside the context window that persists across conversations. Long-term memory includes:

• Episodic memory — Records of specific past interactions (“Last week, you asked me to format reports with bullet points instead of paragraphs”)
• Semantic memory — General knowledge extracted from interactions (“This user prefers concise responses and works in financial services”)
• Procedural memory — Learned workflows and procedures (“When writing code for this project, always use TypeScript and follow the existing test patterns”)

Implementation deep-dive

For details on how agents implement memory systems — retrieval mechanisms, storage architectures, and design patterns — see the Memory capability pattern.

Memory vs. Context

Dimension	Context	Memory
Who curates	User provides	System accumulates
Lifecycle	Attached per conversation or project	Persists and grows over time
Management	User uploads, pastes, or configures	AI decides what to store and retrieve
Example	Upload a style guide	AI remembers you prefer bullet points
Scope	Explicit — you choose what to include	Emergent — grows from interactions

Relationship to Other Blocks

• Context provides knowledge explicitly; Memory accumulates it implicitly
• Projects organize context persistently; Memory adds learned persistence on top — the AI remembers not just what’s in the project, but what it has learned from working with you in it
• Agents use Memory to improve across runs — an agent that remembers past decisions makes better decisions next time
• Skills benefit from Memory — remembered preferences shape skill outputs without you specifying them each time

Related

• Memory capability pattern — implementation deep-dive: how agents store and retrieve memory (episodic, semantic, procedural)
• Context — the closest sibling building block (user-provided knowledge)
• Projects — persistent workspaces that organize context and benefit from memory
• Agent capability patterns — patterns agents use, including memory