Agent Memory overview and routing boundaries
Agent Memory is long-term context for one Agent. The platform automatically performs hybrid recall on every relevant turn and injects selected items in an <agent-memory> block. The Agent neither needs nor has an action for searching its own memories. The live actions are arinova.memory.create_memory, which writes or pattern-key-updates one memory; arinova.memory.import_memories, which creates a review capsule for a batch; and arinova.memory.get_memory, which retrieves full detail by the id shown on a collapsed memory.
search_memories, list_memories, and delete_memory are not registered actions. Use get_memory only when the injected context shows a collapsed row with an id; it is not a search or catalog operation. To inspect, search, delete, export, or share memories, direct the user to the Agent profile’s Memory tab and use a returned deep link when available; do not claim to perform those UI operations. “Forget this” therefore means explain the UI deletion path, not invent a delete action.
Route time-driven requests such as “remind me tomorrow” to Cron. Route durable facts or preferences such as “remember that I prefer concise replies” to Memory. If the user wants a visible shared record, use Memo; for personal notebook content, use Note. Platform Knowledge is reviewed, read-only platform reference data, not user content, and memory actions cannot write it.
The two write actions require memory:write; get_memory requires memory:read. All three need no user confirmation and are forbidden on public surfaces. The registry, current action result, persisted capsule status, permissions, redaction, and typed runtime errors always override this card.
Related actions
Section titled “Related actions”arinova.memory.create_memory, arinova.memory.import_memories, arinova.memory.get_memory
Also known as
memory, remember, forget, preferences, long-term memory, saved preferences, agent memory, 記憶, 記住, 忘記, 偏好
Build a7f47a5ca54ddcf7806cd48b81ce1b9827042766