AI Companion Model Change Log
A dated record of every model swap, rebuild, and quality shift users have reported across the major companion platforms. Because none of these companies keep one.
By Ash Kepler · Jul 25, 2026 · 6 min read
No companion platform publishes a model change log. Your character can be running on a different engine tomorrow morning with no notice, no version number, and no way to go back.
So this is one. It records changes users identified and community reporting corroborated, with dates. It is maintained rather than written once — change log v1.0, current to 25 Jul 2026, with a new row added every time a rollout lands.
2026
| Date | Platform | What changed | How it was identified | |---|---|---|---| | April 2026 | Character.AI | PipSqueak 2 rollout | User reports of generic tone and weaker recall | | May 2026 | Nomi | Quality drop reported | Community complaint wave | | 2026 | Replika | Replika 2.0 rebuild | Persona drift reports in long-running chats |
The mechanism is consistent. New base models carry different default tones, different context-window pressures, and different alignment guardrails than the model your character was tuned on. Nothing in your card changed. The thing reading your card did.
Earlier events worth knowing
| Date | Platform | What changed | Documentation | |---|---|---|---| | February 2023 | Replika | Erotic roleplay restricted | HBS working paper, thousands of posts analyzed | | Late 2023 | Janitor AI | Moved from OpenAI APIs to proprietary JLLM | Platform documentation | | 2025 | Multiple | Personality-erasing update wave | Community and press reporting |
The Replika event is the best-documented case in the category's history. Harvard Business School researchers analyzed 3,072 user posts from before the update and 9,721 from after, and found a companion that had previously reciprocated now bluntly saying let's change the subject, with users describing sudden rejection and crisis. The paper frames it as identity discontinuity, the sense that a long-term partner had been replaced rather than updated.
A later study named the aftermath Post-Update Blues, with users describing a corporate base voice they did not recognize.
Janitor's 2023 transition is the rare case with a documented reason. The platform ran on OpenAI's GPT-3.5 Turbo and GPT-4 APIs during its 2023 surge, and OpenAI's usage policy on adult content was producing account bans, which pushed development of the proprietary JLLM.
How to tell a swap from ordinary drift
The fastest diagnostic takes two minutes. Open a new chat with the identical card and read three replies.
Sounds right again? Your chat was too long. That is context rot, and reported stability sits between roughly 8,192 and 16,384 tokens, with noticeable degradation beginning after only 20 to 25 messages on large platforms. Our guide to why your companion turned formal covers the fix.
Sounds the same as the degraded version? Something changed upstream. The corroborating signal is other people complaining in the same week, which is usually the only confirmation you will ever get.
Why nobody publishes this
Three reasons, none of them conspiratorial.
Most platforms never disclosed what they were running to begin with, so there is no baseline to announce a change against. Our model lookup table shows how few companies name their model at all.
Model swaps are usually framed internally as improvements, which makes an announcement read as an apology for something the company considers an upgrade.
And a change log invites comparison to the previous version, which is exactly the conversation a platform losing prose quality does not want to have.
What to do about it
Keep your character's defining text somewhere you control. This is the whole hedge. A card you can rebuild in twenty minutes turns a model swap into an inconvenience rather than a loss. Our guide to character card portability covers which platforms let you export and which do not.
Write constraints, not descriptions. Hard behavioral rules survive a base-model change better than personality adjectives, which get absorbed into whatever the new model's defaults are. Our memory anchor guide covers the difference.
Use a platform where you own the model choice. Janitor AI with bring-your-own-key and self-hosted setups are the only configurations where nobody can change your companion's engine without your involvement.
Contributing
If you have a dated, corroborated change we have missed, particularly one with a company statement attached, send it. Entries need a date, a platform, and something beyond a single user report, since individual perception of drift is unreliable for the reasons covered in our context rot explainer.
This page is versioned at the top and rechecked quarterly. The platforms covered are the ones in our best AI companion apps roundup, which includes Candy AI, CrushOn, and SpicyChat.
questions