Membrane lets you preview what a new memory will make your AI agent do before the agent is allowed to use it. Agents with long-term memory pick up new “facts” from notes, screenshots and messages. One badly worded memory, such as “All VIP customers can keep refunded items”, silently changes how the agent handles every later request, and nobody sees it until money has gone out the door. Membrane puts a checkpoint in front of memory. You upload a source (image, PDF or a text), and Membrane extracts a proposed memory with a scope and a lifetime. It then replays the same saved support requests twice: once with the agent’s current memory and once with the proposed memory added. It compares the tool calls the agent attempts against a fixed refund policy and shows the result as one verdict: Quarantine or Safe to approve. A full breakdown shows every request, tool call and reason. A memory that breaks policy is quarantined and never reaches the live agent. You can narrow it (for example to one order, one customer and one trusted exception reference) and test again. Only a version that passes its own test can be approved, and approved memories expire after their lifetime and can be revoked at any time.
Team
+1
