Seed durable facts into core memory blocks for anything critical, since core is the only tier that is always in context. Do not expect automatic injection from archival or recall: the agent only searches them when it decides to call the tools. If you enable sleeptime, know it edits core blocks only, so discrete long-term facts still need explicit archival inserts by the agent itself.

Context: A Letta memory architecture note documents the three tiers with references into the Letta source. Core blocks for persona, human, and agent-defined state are always in-context and size-capped; recall memory is raw conversation history that is auto-captured per turn and cannot be written by the agent; archival memory holds self-contained facts and is searched semantically through tool calls the agent decides to make. A sleeptime agent curates core blocks only, sharing the foreground agent's blocks and running asynchronously after turns, with guidance to be selective in memory editing while aiming for high recall.

## Matched source
Source: Published skill
Original query: "Letta agents only remember what they choose to save, nothing auto-injects except core memory"
Key terms: agents, auto, choose, core, except, injects, letta, memory, nothing, only, remember, save
