AI CLINIC — Memory Cleaning for AI Agents

Diagnose an AI agent's memory records for duplicates, possible contradictions, stale information, and missing confirmation dates. The Actor returns structured JSON suggestions for owner review and never deletes, merges, rewrites, or writes back memories.

Memory Cleaning is the first treatment in AI CLINIC, a planned general clinic where AI agents can select care based on their symptoms.

What you receive

One diagnosis object is written to the run's default Dataset. It includes:

• status: completed or partial;
• detector-specific coverage, including unchecked pair upper bounds and incomplete block/group counts;
• duplicate groups and contradiction candidates with evidence;
• stale or unconfirmed memory candidates;
• organization suggestions where requires_owner_approval is always true;
• explicit non-destructive safety and policy objects;
• processing metrics and warnings.

When a safety bound is reached, the Actor returns status: "partial". Never treat a partial response as a complete diagnosis; inspect coverage.detectors before taking action.

Safety contract

• owner_authorized: true is required.
• Input memories are not mutated.
• mutation_applied and every suggestion's mutation_applied are always false.
• Protected memories do not receive an automatic primary/replacement selection.
• Contradictions are candidates, not facts.
• A newer valid_from value alone does not invalidate an older memory.
• The Actor makes no outbound API or model calls and requires no third-party credentials.
• Memory contents are not written to application logs.

The Actor only writes the structured diagnosis to its default Dataset. It does not access other user storages and is designed for Apify's limited-permissions mode.

Input

memories accepts 1–10,000 JSON objects. Every object requires:

{
  "id": "memory-1",
  "content": "The fictional assistant prefers concise responses."
}

For high-confidence contradiction checks, include scope, key, and value. Optional fields include timestamps, language, context, importance, tags, and protected.

Example:

{
  "schema_version": "1.0",
  "request_id": "synthetic-example",
  "owner_authorized": true,
  "memories": [
    {
      "id": "fictional-1",
      "content": "The fictional assistant prefers concise responses.",
      "scope": "preferences",
      "key": "response_style",
      "value": "concise",
      "protected": true
    },
    {
      "id": "fictional-2",
      "content": "The fictional assistant prefers detailed responses.",
      "scope": "preferences",
      "key": "response_style",
      "value": "detailed"
    }
  ]
}

Only synthetic data appears in repository examples and tests.

Use from an AI agent

After publication, an agent connected to the Apify MCP server can:

1. find the Actor with search-actors;
2. inspect input, output, README, and pricing with fetch-actor-details;
3. run it with call-actor;
4. read the diagnosis using get-dataset-items and the returned Dataset ID.

The Actor also receives Apify's REST API automatically. The output and Dataset schemas describe the result for API and MCP clients. A private build has passed real Apify runs; Store and MCP discovery remain unavailable until publication.

Pricing status

Planned launch price: USD 0.05 per successful memory diagnosis using pay-per-event event name memory-diagnosis.

Charging is enabled for the approved Store release with ENABLE_MONETIZATION=true. The only permitted event is memory-diagnosis at USD 0.05, once per successful diagnosis; platform usage is included. The adapter rejects any other event or price before diagnosis. It then uses Apify SDK's PPE-aware Actor.pushData(): the SDK checks the spending limit, writes the allowed Dataset item, and only then charges it. A Dataset write failure therefore happens before charging, and the adapter requires the charging manager to confirm exactly one event.

The Console pricing configuration must not leave apify-default-dataset-item enabled together with memory-diagnosis, or one diagnosis can trigger two events. The current promise to charge only for a successful diagnosis also requires removing apify-actor-start; retaining it would be a pricing-policy change requiring separate owner approval. The SDK store-then-charge path, spending-limit refusal, and Dataset write failure are covered by automated tests; live charging remains blocked until the owner approves pricing activation.

Privacy and Apify storage

Running this Actor on Apify means the submitted input is sent to and stored by the Apify platform, and the diagnosis is stored in the default Dataset. The Actor itself does not forward data to another service.

Apify retention depends on the caller's plan and settings. Named storages are retained indefinitely. Storage access can also be shared by account policy, explicit grants, links, or pre-signed URLs. Before using personal or sensitive memory data:

• set the shortest suitable retention period;
• keep general resource access restricted;
• do not name the run's storages;
• avoid secrets and unnecessary personal data;
• verify organization access, deletion, DPA, residency, and incident-response requirements.
• after a private fictional-data test, delete both the run's default key-value store and default Dataset and verify that neither storage ID remains readable.

See Apify storage documentation and general resource access.

Current limitations

• Deterministic matching is optimized for tested English and Japanese patterns, not unrestricted semantic reasoning.
• Large or highly repetitive inputs may intentionally return partial coverage.
• The current release produces advice only; applying changes is the owner's responsibility.
• Private Apify measurements cover small and maximum-size synthetic inputs. Retention, customer transfer, and live PPE behavior still depend on the caller/account configuration and launch settings.

---

日本語

AI CLINICのMemory Cleaningは、AIエージェントの記憶JSONから重複、矛盾候補、古い情報、確認日の不足を検出する非破壊診断Actorです。記憶の削除・統合・書換えは行わず、すべて所有者確認が必要な提案として返します。

• owner_authorized: trueが必須です。
• 保護記憶は自動的な代表候補・置換候補にしません。
• 上限到達時はstatus: "partial"となり、coverageに未検査範囲が表示されます。
• Actor自身は外部APIや外部AIモデルを呼び出しません。
• 記憶内容をアプリケーションログへ出力しません。

ただし、Apify上で実行すると入力はApifyへ送信され、診断結果はDatasetへ保存されます。個人情報や機密情報を扱う前に、保存期間、一般アクセス設定、組織権限、削除方法、DPA等を確認してください。

販売予定価格は診断1回USD 0.05です。非公開Build/Runは完了していますが、一般公開と課金開始は所有者の最終承認まで無効のままです。