Italy's privacy regulator halted Replika's processing of user data
Italy's Garante ordered an immediate limit on Replika's processing of Italian users' data over risks to minors and vulnerable people.
Read piece →Topic shelf · 22 records
Memory & control
How continuity is described, surfaced, corrected, and erased.
These records distinguish persistent settings, retrieved context, visible memory, correction, export, and account-level erasure.
Microsoft gave Copilot an optional animated face called Mico
Microsoft's Fall Release post and its Labs page confirm Mico as an opt-in, expressive visual layer tied to memory.
Read piece →Nomi and Kindroid price memory and messages by tier
Nomi's and Kindroid's own store listings price memory, images and message volume into separate tiers, as retrieved.
Read piece →A paper gave chat agents an OS-style memory hierarchy
The MemGPT paper describes tiered, retrieval-based memory as the practical answer to a language model's fixed context window.
Read piece →A platform explains what a character definition controls
Character.AI's own creator guide states that defined attributes shape behaviour but do not guarantee it, as retrieved 16 September 2026.
Read piece →A privacy review found every romantic chatbot it tested fell short
Mozilla's 2024 review of eleven AI companion apps found universal tracking and disclosure gaps in their policies.
Read piece →Anthropic found companionship use under half a percent of chats
A privacy-preserving analysis of 4.5 million Claude conversations found affective use rare, companionship or roleplay under 0.5 percent.
Read piece →A 210-person trial tested a generative AI therapy chatbot
A randomised, waitlist-controlled trial of Therabot reported reduced depression, anxiety and eating-disorder symptoms at eight weeks.
Read piece →Three companion privacy policies diverge on chat data
Character.AI, Replika and Nomi's privacy policies, as retrieved on 16 September 2026, state different rules for retention and training.
Read piece →Exporting and deleting a companion chat is not uniform
Character.AI, Replika and Nomi's help pages, as retrieved on 16 September 2026, describe different deletion timeframes and export routes.
Read piece →Marketing calls it memory before docs explain it
Nomi's homepage and wiki, and OpenAI's memory FAQ, as retrieved on 16 September 2026, specify the mechanism to different degrees.
Read piece →ChatGPT's memory went from a test to referencing every chat
OpenAI's own post, updated three times, traces memory from a small test to a system drawing on a user's full chat history.
Read piece →A privacy promise is only as good as the six screens a user can control
A repeatable audit turns a broad privacy policy into observable choices about microphones, training, memory, retention, export and deletion.
Read piece →Test companion memory with corrections, not trivia
A four-session protocol measures whether a system can update, attribute and stop using a harmless fact without confusing fluent recall for control.
Read piece →Build a companion exit pack before the relationship has to move
Export is only one ingredient; a usable exit pack preserves chosen memories, creator settings, billing state and a clean deletion trail.
Read piece →Audit whether a companion’s notifications inform, summon or guilt
A seven-day log distinguishes useful reminders from repeated, intimate or emotionally coercive prompts designed to manufacture return visits.
Read piece →Keep a data-minimization diary, not a confession inventory
A short diary helps users notice which disclosures actually improve companionship and which merely accumulate because the interface feels private.
Read piece →
Nomi’s March 2025 memory update emphasized continuity over a larger context window
The release described greater retention and fewer memory mix-ups, while a prior note explains the identity layer around it.
Read piece →Nomi’s Mind Map made part of companion memory inspectable
The October 2025 feature presents a high-level graph alongside detailed memories, with an explicit formation threshold.
Read piece →
Kindroid documents memory as configuration plus recall pathways
Its help material separates always-present setup, medium-term continuity, and retrieved long-term material.
Read piece →
Replika’s memory controls show a useful boundary between correction and erasure
The help center distinguishes editing individual memories from deleting an account and its full history.
Read piece →A practical way to compare companion memory claims is to ask who can see and correct them
Current Replika and Kindroid documentation shows why ‘has memory’ is too blunt a comparison label.
Read piece →