Memwork
Back to blog

Black-box memory vs. a memory you own

How today’s AI assistant memory works (using Grok as a documented example) and why Memwork keeps agent memory as readable docs you own, see and edit.

Oct 7, 2026MemworkMemwork

AI assistants increasingly "remember" you. The open question is who can see that memory, and who controls it. This post compares a typical black-box approach with Memwork, a shared memory space for many agents.

How black-box memory works today

We use Grok as the main example because xAI publishes its system prompts, which makes its behavior verifiable. xAI added memory to Grok in April 2025 (TechCrunch). Based on its public docs and published system prompt (October 2026):

  • Memory is tied to conversations. What Grok remembers comes from your past chats.
  • The mechanism is not public. How memories are stored and retrieved is not documented.
  • No per-item view or edit. There is no list of individual memories to inspect or correct. To make it forget, you delete the conversation or use the book icon.
  • No confirmation of forgetting. The published system prompt tells the model never to confirm to the user that it has forgotten a memory.

Other assistants take different approaches. Some, such as ChatGPT's saved memories, offer a management screen. Still, in most products the memory lives inside one vendor's assistant and is not available to your other tools and agents.

Memwork: a white box

Memwork treats memory as content you own:

  • A shared memory space for many agents. One place that all your agents use, instead of one silo per assistant.
  • You own, see and edit every memory. Each memory is a readable doc. You can open it, fix it, or delete it, and nothing is hidden.
  • Distilled into docs. What matters in your bot and agent interactions becomes readable documents you own.
  • Open access via MCP. Multiple agents read and write through MCP, always under your own permissions.

At a glance

| | Black-box memory (e.g. Grok) | Memwork | | --- | --- | --- | | Scope | Tied to conversations | Shared across many agents | | Mechanism | Not public | Readable docs | | View / edit items | No per-item list | Every memory visible and editable | | Forgetting | Delete the conversation or use the book icon | Edit or delete the doc | | Agent access | Inside one assistant | Any agent via MCP, under your permissions |

Sources

Grok is a trademark of xAI. It is named here only as a factual reference. Comparison based on public docs and the published system prompt as of October 2026.