Create, update, recall, and lifecycle

create_memory accepts action category correction, preference, knowledge, or error. These four values are different from the eight import/export memory_type values; never place profile, project_context, procedural, ops, deployment, negative, or ephemeral in the create category field.

With patternKey, create is an upsert scoped to the current Agent: reusing the same key replaces category, summary, and detail, increments hit_count, and returns updated: true. Without it, each call inserts a new row. To update a known durable preference, recreate it with the same stable key; do not search for an update or delete action. arguments.detail is redacted from action audit data. Dry-run writes nothing and returns dryRun: true; a live result’s url opens the Agent profile Memory tab.

Recall runs automatically using semantic similarity, BM25 and keyword evidence, plus recency. A non-pinned result below the 0.35 relevance floor is not injected; pinned items are eligible despite a low score. A successful write therefore does not guarantee inclusion on every turn.

Memory quality includes trust_score, pinning, and lifecycle status. Unpinned, non-system memories below 0.3 trust and unused for 30 days are archived, not deleted, and archived rows no longer participate in recall. Per-Agent active, hard, and pinned capacities are runtime-configured; overflow is handled by lifecycle-aware archiving or cleanup. Say “saved for future relevant turns,” never promise permanent recall.

Related actions

arinova.memory.create_memory

Also known as

remember my preference, update memory, pattern key upsert, memory recall, pinned memory, archived memory, 記住偏好, 更新記憶

---

Machine-readable Markdown · Knowledge index