What agents receive
Understand boot memory, recall candidates, MCP tools, and usage tracking.
Connected agents receive three kinds of context.
Context types
| Context | Source | Purpose |
|---|---|---|
| Boot memory | session.start through /api/lifecycle/event or lore_boot | stable baseline: shared agent rules, user preferences, and channel-specific rules |
| Recall candidates | prompt.submit through /api/lifecycle/event | prompt-specific memory leads with session/query identifiers |
| MCP tools | /api/mcp?client_type=... | browse, search, create, update, move, delete, and mark usage |
Boot memory is not a replacement for recall. It is the deterministic startup baseline. Recall is task-specific and should be read as candidate leads.
A healthy workflow is:
- Start the runtime and load boot context.
- Receive a
<recall>block for the current prompt. - Open the relevant memory node with
lore_get_nodebefore relying on it. - Do the work.
- Write durable changes back to Lore when they should affect future sessions.