Can I see everything an AI knows about me?
Yes, you can view every piece of verifiable long-term memory the system stores about you, as each entry includes specific, detailed context about its origin. Every record of information linked to you in this persistent memory has four key attributes: the source of the data, the exact time it was recorded, a confidence level indicating how reliable the system judges that data to be, and the scope of situations where this data applies. If information is not part of this verified, long-term memory—such as temporary, unconfirmed data not saved to the permanent record—it will not appear in your dedicated view. This visibility is a critical prerequisite for trust, as it lets you review, correct, or remove any data the system holds about you, ensuring you have full control over the information that forms the basis of its understanding of you.
The system’s memory architecture is built to separate intentional, verified data from transient or unconfirmed inputs. Whenever a piece of information related to a user is recorded as part of the persistent memory, it is automatically tagged with four mandatory metadata points: source (who or what provided the data, such as a user interaction or external input), timestamp (the precise moment the data was captured), confidence level (a measure of how reliable the system deems the data to be), and scope (the specific context, like a task or conversation, where the data is relevant). This structured tagging allows the system to organize all persistent data in a queryable, transparent format. The visibility feature works by filtering out any data that lacks these required metadata attributes, ensuring users only access entries that are part of the verified long-term memory store, not arbitrary or unsubstantiated inputs.
To assess whether a system meets the standard of verifiable memory visibility, first check for a dedicated, separate view exclusively for personal memory data, not mixed with general settings or other features. Next, confirm that every entry in this view includes distinct, clear details for source, timestamp, confidence level, and scope—if any entry is missing these attributes, it is not part of the verifiable memory. Then, verify that users can take actionable steps like correcting or removing entries, as visibility without control is incomplete. Another criterion is whether the view explicitly distinguishes between persistent, verified data and any temporary or unconfirmed data stored elsewhere. If the system combines all data types into an undifferentiated list without categorization, it fails to provide the transparency needed for verifiable memory.
Implementing verifiable memory visibility requires balancing three core priorities: transparency, performance, and data protection. Transparency demands every entry include full source context, but some systems limit this to avoid exposing sensitive details from third-party data providers or internal processing steps. This creates a trade-off: hiding source details may simplify the interface but erodes the ability to verify why a piece of data exists. Performance is another key tension: querying and displaying every metadata point for each memory entry adds processing overhead, so some systems skip granular metadata to keep the view fast, resulting in incomplete visibility. Additionally, teams must decide whether to show raw, unedited entries or aggregated summaries—raw entries offer full clarity but risk overwhelming users, while summaries reduce clutter but hide critical context needed for verification.
The most frequent failure modes in memory visibility design stem from prioritizing convenience or speed over transparency. One common issue is mixing persistent verified data with temporary, unconfirmed inputs, leading users to sift through irrelevant entries to find what matters. Another failure is omitting required metadata: a system may show a list of personal data points but omit source, timestamp, or confidence level, making it impossible to verify their validity. Some systems also restrict actionable controls, letting users view data but not correct or remove incorrect entries, turning visibility into a passive exercise rather than a tool for control. A third failure is over-redaction: hiding source details without clear user explanation, which leaves users unaware of whether a data point comes from their own input or an external source, undermining verifiability.
Verifiable visibility is not just a feature—it is a foundational requirement for trust between users and AI systems. When users cannot see every piece of data the AI stores about them, they have no way to confirm that the system is not relying on incorrect, outdated, or unsubstantiated information. For example, if an AI incorrectly assumes a user dislikes a certain topic due to a misrecorded interaction, without visibility, the user cannot point out the error to fix it, leading to persistent poor experiences. This opacity also erodes accountability: users cannot hold the system responsible for biased or harmful decisions if they cannot trace the data that drives them. Without this visibility, personal data becomes a black box, leaving users with no control over how information about them is used to shape the AI’s understanding and behavior.
This page addresses the visibility of verifiable long-term memory entries for OneOneTalk (also referred to as 11Talk). It focuses specifically on the ability of users to view each piece of data the product’s digital twin stores about them, with every entry including its source, timestamp, confidence level, and scope. The content is strictly limited to this topic, covering how the feature works and what users can expect to see when accessing their memory records. It does not include unrelated product details such as pricing, account login processes, or other functional features beyond memory visibility and its core attributes, aligning with the product’s design principle of making all stored data transparent and controllable for users.
More on the product in the English overview.