Memory Block
A memory block is a labeled, persistent chunk of agent memory directly editable by the agent via tool calls. The term originates with Letta (formerly MemGPT). Conventional blocks include `human` (what the agent knows about the user) and `persona` (the agent's own self-description), with custom blocks for project context, preferences, or domain-specific state.
Memory blocks live in main context — the agent sees them on every turn — and are edited via tools like `core_memory_append` and `core_memory_replace`. The structure makes semantic memory programmable in a way unstructured prompt-stuffing does not.
Example
A coding-assistant agent has memory blocks named `user_preferences` (preferred languages, code style) and `project_context` (current repo, recent files touched). When the user mentions "I'm switching to TypeScript for this project," the agent calls `core_memory_append` on `user_preferences` to record the change — and from the next turn onward, every prompt automatically includes that fact.
Frequently asked questions
What is Memory Block?
- A memory block is a labeled, persistent chunk of agent memory directly editable by the agent via tool calls. The term originates with Letta (formerly MemGPT).
How does Memory Block work?
- Memory blocks live in main context — the agent sees them on every turn — and are edited via tools like `core_memory_append` and `core_memory_replace`. The structure makes semantic memory programmable in a way unstructured prompt-stuffing does not.
Can you give an example of Memory Block?
- A coding-assistant agent has memory blocks named `user_preferences` (preferred languages, code style) and `project_context` (current repo, recent files touched). When the user mentions "I'm switching to TypeScript for this project," the agent calls `core_memory_append` on `user_preferences` to record the change — and from the next turn onward, every prompt automatically includes that fact.