Skip to content
AX Principles
AX-SYS-33Intermediate

Scope agent memory, and let people see and erase it

Delimita la memoria del agente y deja que las personas la vean y la borren

#all-industries#saas

Store what an agent remembers in a place you control, scoped to one user and one project, with an expiry, and give people a way to read, correct and delete each entry.

Why

Memory lets an agent pick up where the last conversation ended. It also turns one bad inference into a standing instruction: if the agent wrongly records that a customer prefers store credit, every later refund follows that note, and nobody can fix what nobody can see. Anthropic's memory tool shows where the responsibility sits. It operates client-side: the model only requests file operations, and your application decides where and how the data is stored, for example in a per-user directory or as keys in a database. Its security guidance also falls to you: strip sensitive data before a write, cap how large a file can grow, and periodically delete memory files that have not been accessed in a long time. Anthropic applies the same rules to people in its own product. In Claude, a user can see exactly what is remembered, open any topic, edit or delete it, keep each project's memory separate from other projects, and reset all of it permanently. A product that keeps memory for agents owes its users the same controls. AX-SYS-16 keeps one task's progress alive; this principle covers what persists between tasks and who can inspect it.

Do

Key memory by user and project, expire entries nobody uses, and expose list, edit and delete for every entry and for the whole store.

Don't

Keep agent memory in a store only the agent can read, or let it hold secrets and stale facts indefinitely.

Artifact

GET /v1/agents/acme-support-bot/memory?user=usr_481&project=refunds
{
  "scope": { "user": "usr_481", "project": "refunds" },
  "entries": [
    {
      "id": "mem_07",
      "text": "Prefers refunds to the original card, not store credit",
      "source": "conv_2291",
      "last_used": "2026-09-28",
      "expires_at": "2027-03-28",
      "actions": ["edit", "delete"]
    }
  ],
  "delete_all": "DELETE /v1/agents/acme-support-bot/memory?user=usr_481&project=refunds"
}

Each entry names the conversation it came from, so a person who finds a wrong fact can see where it started before deleting it.

Citations

  1. [01]
    Claude Docs: Memory tool ↗

    The memory tool operates client-side: Claude requests file operations and your application executes them, so you control where and how the data is stored, for example in a per-user directory or keys in a database. Its security guidance says to strip sensitive data before writing, cap file sizes, and periodically delete memory files that haven't been accessed in a long time

  2. [02]
    Claude Help Center: Use Claude's chat search and memory to build on previous context ↗

    Users can see exactly what Claude remembers in Settings > Memory, select any topic to read it, edit it or delete it; each project has its own separate memory space; Reset memory permanently deletes all memories, including project memories