Memory & sessions
DOCS / WORK WITH VIBES

Memory & sessions

Stay oriented across a long task without carrying every detail into every prompt.

Each session has its own agent

The runtime owns session lifecycle and caches one agent instance per session. The API’s stream coordinator owns transport state such as abort controllers and reconnect tracking.

Session data persists in SQLite through the storage backend. The repository includes a Drizzle backend and configuration for different storage environments.

A scratchpad and a note store

A per-session scratchpad keeps immediate working context. Long-term notes are searchable through remember, recall, update_memory, forget, and list_memories.

The prompt receives a compact index rather than the full note store. This keeps memory useful without making every model call carry all historical text.

State you can inspect

text
scratchpad.md
plan.md
tasks.json
tracked_files.json
artifacts/
subagent_results/

These session files capture working context and outputs. The shared note store is separate from per-session state.

Context management

Restorable compression swaps large content for file or path references. Rolling summaries help fit long conversations into the context window. Errors are preserved separately, and stable prompt prefixes support model-provider caching.

Improve this page Grounded in the Vibes source.

Find a guide, command, or concept.