The recent JanitorAI memory feature glitch is the most discussed topic in the roleplay community these past few days.
If you felt your long-term JanitorAI character had a “personality shift” last week—forgetting your agreements, becoming vague, or failing to recall the last scene—don’t rush to delete and restart; you aren’t the only one. Between August 18th and 20th, an experimental memory feature intended only for a small test group was accidentally released to “all users,” causing memory issues across a vast number of conversations. The official team confirmed that nearly 15,000 chats were affected by the secondary bug where “deleting a message cleared the memory.” This article breaks down the full timeline, how the team apologized and recovered the data, and what this means for players currently using JanitorAI.
Understand: A Core Feature Glitch on a 15M+ User Platform
JanitorAI’s homepage slogan is “Be anyone. Build anything.” and explicitly states the platform has “15M+ users crafting living stories.” For readers who have long followed AI roleplay, JanitorAI has been one of the primary platforms in this space, accumulating high activity through open third-party AI model integration and a massive library of user-created characters.
On a platform of this scale, “Chat Memory” is not an optional tool; it is a core feature that determines the quality of the roleplay experience—it summarizes past interactions so the character remembers what happened between you, even after hundreds of messages. When this feature fails, it doesn’t just affect a niche function; it impacts “every ongoing character relationship.” This is why the mid-August incident caused such a massive reaction in the JanitorAI player community.
Full Timeline: Official Records from 8/18 to 8/24
The JanitorAI team maintains a public Changelog page, known for its casual, non-corporate tone, announcing system changes in real-time, including glitches and fixes. From this public record, we can piece together the entire event:
- August 18: An experimental “memory replaces old messages” feature, originally locked to a small test group, was accidentally released to all users. For chats with stored memory summaries, the messages covered by the memory were no longer sent to the model, meaning the character had less context, leading to vague replies and missed details.
- August 19: The team issued an apology, admitting the test “escaped the original test group and affected everyone,” and detailed 5 technical points involved. Simultaneously, the “memory replaces old messages” feature was disabled and changed to an opt-in setting.
- August 20: The team provided an update, identifying the “last remaining experimental residue”: if a player deleted a message covered by memory, the memory itself would be cleared. This secondary bug affected approximately 15,000 chats between August 18th and 20th. The good news is that the team confirmed this bug only “hid the memory rather than destroying it,” so most affected memories were successfully restored, except for content overwritten after August 18th.
- August 24: An unrelated incident—the character search service briefly malfunctioned, which the team announced they were investigating. This was unrelated to the memory feature but showed the system was experiencing significant turbulence during this period.
The screenshot above is the original August 20th update from the JanitorAI Changelog—the language is candid and informal, which is quite rare in corporate communications.
Official Breakdown: Where Exactly Did It Go Wrong?
More important than the fact that a glitch occurred is that in their August 19th explanation, JanitorAI rarely broke down the incident into 5 specific technical points:
- Memory replacing old messages: This was the starting point. The intent was to save resources by not re-sending summarized content to the model, but this mechanism was accidentally released to everyone, reducing the actual context the character could see.
- Summary failure clearing memory: If the automatic summary generation failed, was canceled, or resulted in empty content, the system would directly overwrite the player’s existing memory field with empty data.
- Model “internal thoughts” leaking into memory: Some models generate a draft or thought process before replying. In some cases, this text was misidentified as official content and written directly into the memory field, polluting the summary.
- No safe update method for manual memory: If players manually edited memory, there should have been an option to “keep old content and only update new parts.” This was hidden during the accident, leaving players with only an “overwrite all” button, increasing risk.
- Automatic summary feature was part of the same experiment: Because automatic summarization was bundled with the “memory replaces old messages” experiment, the impact was much broader than a single feature.
The team explicitly stated that all 5 points have been fixed: manual memory now has a “keep updates” option; model thought text is filtered out on both web and app; and a failed summary generation will no longer wipe the existing memory display.
The Most Dramatic Twist: Deleting a Message Wiped the Memory
Compared to the August 19th breakdown, the August 20th update was more alarming—it revealed an unexpected dependency: deleting an old message that happened to be covered by the memory summary would also clear the memory itself.
This was not a rare edge case. Players delete messages for many reasons: to redirect the plot, remove a hallucinated AI response, or simply clean up chat logs—no one expected this action to be linked to memory loss. The team admitted this bug affected about 15,000 chats in two days.
It is worth noting the team’s handling of this: after confirming the bug “hid” rather than “destroyed” the memory, they restored most of it, with only the small portion overwritten after August 18th being unrecoverable. The team also clarified two things: the branching feature never touched anyone’s memory; if memory appeared empty after branching, it was simply because the original chat’s memory had already been cleared by the delete bug. Additionally, the “context exceeds the limit” error that caused summary failures for long memories has also been fixed.
If You Are a JanitorAI Player, What Should You Do Now?
The incident is resolved, but if you are a long-term JanitorAI player, a few things are worth checking. First, go to the Chat Memory panel (on both web and app) and check the status of the “Memory replaces old messages” toggle—it is now off by default, reverting to the original behavior where memory is layered over the chat without deleting anything, unless you manually turn it on. Second, if you experienced sudden character amnesia or blank memory fields between August 18th and 20th, check if the memory has been restored—the team has handled most cases, but if not, it may be the small portion that was overwritten after August 18th. Third, this new toggle currently only works on the web version; the app will sync in the next update. If you primarily use the mobile app, it is normal not to see this option yet.
It is worth mentioning that just days after the memory incident concluded, JanitorAI launched the “Similar characters” feature on August 28th, allowing players to see more similar characters directly on the character page—showing that product development did not stop while the team handled the aftermath. If you are interested in other recent JanitorAI changes, such as the janitor+ Router which allows direct connection to 70+ AI models without a proxy, you can review that as well.
Not an Exception: Why AI Companion “Memory” Always Fails First
If you have followed this field long enough, you will find that “memory” is the most prone-to-failure feature in all AI roleplay and companion platforms; JanitorAI is not the only case.
Third-party review site RoboRhythms pointed out in their JanitorAI memory guide that the most common technical reason for memory failure is the token count exceeding the model’s context window—the window for the free JLLM model fluctuates between 5,000 and 9,000 tokens depending on actual conditions. Once the chat length exceeds this, the earliest messages are automatically pushed out, and the character naturally “forgets” the initial details. In other words, even without this experimental bug, the memory feature itself has structural fragility in long conversations.
Another third-party site, LumiChat AI, positions Chat Memory in their guide as a concise, editable summary of the session rather than a full transcript, suggesting players manually update important relationship details, current scenes, and confirmed facts regularly, rather than relying entirely on automatic summaries. This mindset of backing up important settings yourself is a reminder to all AI roleplay players from this incident.
Looking at the bigger picture, this is a common challenge for the entire industry. Character.AI recently added a three-layer memory architecture with Story Memory and Facts; Kindroid, after transitioning to V2, focuses on “long-term memory” and location awareness, but simultaneously slashed free usage quotas; as for the entire AI girlfriend/companion app sector, “remembering you” has been treated as a core competitive advantage more important than visuals for years. As memory features become more complex, more points of failure emerge. JanitorAI’s experience is a concrete demonstration of the shared risks in the industry.
If You Are Looking for a Platform with Simpler, More Transparent Memory
After reading about this incident, if you are considering other options, AI Zhimeng is worth considering. AI Zhimeng focuses on anime-style immersive interactive fiction and dual-mode roleplay, with an interface fully supporting Traditional Chinese, making it more intuitive for players in Taiwan and Hong Kong.
Compared to tracking when various experimental features might quietly go live on existing platforms, AI Zhimeng’s current conversation and character setting mechanisms are simpler and more transparent. If you are a creator, you can also earn a 30% point rebate by sharing your designed characters; the platform offers +500 points for daily check-ins, and you can get 15% off your first top-up by entering the code WELCOME15. Players looking for a new environment to manage their characters might want to take a look.
FAQ
Q1: Is my JanitorAI memory really recovered?
If your chat was affected by the “deleting a message cleared the memory” bug between August 18th and 20th, the team states that most have been restored. You can check directly in the Chat Memory panel of that conversation. The only unrecoverable part is the small portion overwritten by new content after August 18th, which the team admits cannot be restored from the system side.
Q2: Will the “Memory replaces old messages” feature appear again?
Yes, but it is now an opt-in setting that needs to be manually enabled in the Chat Memory panel. This means unless you choose to enable it, your memory behavior will remain in the original mode of layering over the chat without deleting anything.
Q3: Will this incident make people afraid to use JanitorAI’s memory feature in the future?
This concern is understandable, but from another perspective, the team’s choice to publicly and detailedly break down the 5 points of failure, rather than glossing over them, shows that the issues were taken seriously and fixed. A more practical approach is to develop the habit of backing up important character settings and key plot points yourself, rather than relying 100% on any platform’s automatic memory system—this applies to all AI roleplay platforms, not just JanitorAI.
Q4: If I am an iOS or Android App user, do I need to do anything?
Currently, this new “Memory replaces old messages” toggle only works on the web version; the team stated it will sync in the next app update. If you primarily use the mobile app, it is normal not to see this option for now, and the memory behavior itself has already reverted to the original default state, so no extra action is needed.
Conclusion
An experiment that should have only affected a small group of test players turned into a storm affecting tens of thousands of chats due to an accidental full release. JanitorAI’s handling of this was not perfect—releasing outside the test group was a mistake—but the subsequent days of public, detailed, and candid explanations and recovery efforts at least showed a commitment to player relationships. If you are a heavy roleplay player, this incident may be a reminder: developing the habit of keeping an extra backup of your most cherished character settings and plot memories is more reassuring than trusting any platform’s automatic memory system completely.
