Built the world’s first verifiable AI unlearning protocol. Solving the $100B compliance gap AI created. Forg3t is to AI what HTTPS is to the Web.
- Tweets 527
- Following 126
- Followers 6,135
- Likes 299
We heard you, Sam.
No worries, we got this.
The era of provable forgetting just began.
@sama
1
2
10
1,605
Privacy controls that run after a prompt is sent are already late. Shield moves the decision to the moment before disclosure, where masking or blocking can still prevent exposure.
75
5/10 @Forg3tProtocol is building the infrastructure for AI to forget, giving companies a way to unlearn specific information, test the result, and prove what happened.
forg3t.io/
128
Hosted APIs, retrieval systems, adapters, and model weights do not share one deletion method. Protocol selects the lane that matches the access level, then records what changed and what could not be reached.
86
One document can travel through a workspace, a retrieval pipeline, an assistant, a cache, and a vendor API. Trace maps the path so the response matches the systems that actually received it.
65
A browser prompt can contain customer data, credentials, internal strategy, or source code. Shield evaluates the risk before send, without keeping the prompt itself. Control begins one click earlier.
91
A deletion request needs a defined target, a system map, the right intervention lane, before-and-after tests, and an evidence bundle. Skip one step and “deleted” becomes an assumption.
1
107
AI data exposure is not a yes-or-no question. Trace separates what was observed, declared, inferred, and still unknown. Good governance starts by showing uncertainty instead of hiding it.
86
The safest sensitive prompt is the one that never leaves the device. Shield checks text before it reaches an AI tool, then masks, warns, or blocks according to policy. Prevention belongs at the keyboard.
1
105
Model retention is not one thing.
Information may persist through model weights, adapters, retrieval systems, prompts, caches, logs, or application state.
Before removing anything, an enterprise needs to know which layer it is actually dealing with.
130
Can your AI system forget?
Can you locate every copy and derived artifact?
Can you define the deletion target precisely?
Can you test before and after behavior?
Can an independent reviewer inspect the evidence?
Four “yes” answers are harder than they look.
1
116
AI deletion is rarely owned by one team.
Legal defines the obligation.
Data teams remove records.
ML teams inspect model behavior.
Security reviews access.
Audit asks for evidence.
Fragmented ownership is part of the problem.
1
133
Scenario: An employee leaves a company.
Their sensitive documents are removed from storage, but an internal assistant was previously exposed to them.
The real task is not closing the account.
It is finding every remaining influence path and testing the outcome.
3
173
AI forgetting is not one technical path.
Hosted APIs require behavioral controls and validation.
RAG systems require retrieval revocation.
Self hosted models allow deeper artifact level intervention.
The workflow must match the access level.
2
142
“We deleted it” is a statement.
An evidence workflow records the target, scope, system version, intervention, tests, results, approvals, and delivery artifacts.
Enterprise trust requires inspectable work, not a sentence in a policy document.
3
177
The AI industry has spent years optimizing memory.
The next infrastructure category will be about control:
What entered the system.
What it still retains.
What was removed.
What can be proven.
2
107
A Forg3t workflow begins with five questions:
What must be removed?
Where can it persist?
Which intervention path fits the system?
How will the result be tested?
What evidence must be delivered?
2
121
Removing a document from storage may not remove every derived artifact.
Enterprise RAG systems can include chunks, embeddings, indexes, caches, logs, and generated summaries.
A deletion workflow is only as complete as the system map behind it.
1
133
Your enterprise receives a deletion request.
The record is removed from the database.
What happens next?
A. Nothing
B. The RAG index is updated
C. The model is tested
D. A complete evidence workflow begins
Most companies still do not have one answer.
4
114
AI unlearning is not “making the model refuse a question.”
It is a scoped process for reducing targeted retained influence, validating the outcome, and recording what was done.
2
149