Core Concepts
Memories
Memories store durable guidance that can be re-injected into future requests without retyping the same context.
What memories are
Memories are persisted notes that the runtime can include in prompt assembly on later turns.
- Global memories: available across projects.
- Project memories: scoped to one project path.
- Both are sorted by recency and included when relevant.
How memories affect requests
During prompt building, the app injects global memories and project-local memories into the request context.
- This reduces repeated prompting for stable preferences and constraints.
- Memory content becomes part of model input, so it contributes to context usage and cost.
- Outdated or overly verbose memories can dilute attention just like oversized prompts.
Memory lifecycle
- Memories are created/deleted through runtime storage operations during agent workflows.
- Memory state is persisted and hydrated with the rest of app data.
- Keep memory entries short, durable, and high-signal.
Best practices
- Store stable constraints (coding style, preferred stack, recurring project rules).
- Avoid storing temporary task chatter that belongs in the active session only.
- Periodically remove stale memories so long-term context stays relevant.