Just noticed that the Digital Humans are unable to access their Long-Term Memory. I’m using plugin version 3.6.8 beta 3.3. The digital humans are unable to remember any previous discussions and the Enable Long Term Memory is switched on.
Tried accessing the Dashboard for the Character and the new playground seems to be playing up, and not showing all the memory logs, but when refreshing the browser it shows up.
Any clue what’s going on? I’m still testing out the new plugin V4 with Unreal Engine and have not put it into production.
Any help would be most appreciated.
Character IDs:
60e087de-76af-11ef-80ce-42010a7be011
81b2b358-6cca-11f0-9210-42010a7be01f
The Long Term Memory doesn’t seem to be working. I thought it did yesterday when I tested but this morning, it is not. I am en route to present to a delegation of researchers, industry leaders and academics from Romania.
Any thing you and the team can do please regarding the Long Term Memory and the Error Code 0?
Long Term Memory is scoped to a specific character_id + speaker/end_user profile. So if a request uses a different speaker or end user identity, the character will appear to have a separate or empty memory state.
To use the existing memories, please send:
e44dcc5e-a326-11f0-8464-42010a7be025
Use the same speaker id as:
speaker_id if you are using the legacy gRPC service.
end_user_id if you are using the WebRTC service.
Also, please note that Long Term Memory is not supported for realtime services like openai realtime and gemini live
Thank you. The speaker ID was set up and not been changed after the Unreal Engine project was packaged. I don’t believe it can be changed after a project has been packaged.
Also in the plugin configuration in Unreal Engine when setting up speaker IDs, the screen doesn’t shore a list of the speaker IDs.
How do I fix this issue but reconnecting the original speaker ID to LightArch and Annelore in the packaged project again?
We’re using Claude 4.5 Opus for the characters. Do I need to re package the project?
How can I reconnect the Speaker ID?
The documentation only explains at a surface level and offers no clear instructions. I encourage you to check the setting Speaker IDs with plugin version 3.6.8 beta 3.3 to get a better understanding.
Can you please help get this fixed. Also we need to find a way for the speaker ID to remain persistent.
Everything was working as intended till two weeks back and I’m confused why this would change in its behaviour suddenly.
Happy to hop on a video call to further investigate this with you.
I’ve confirmed in the actual project the user ID is the same for both characters. Yet, it doesn’t seem to be working. New Memory is not being formed and retained beyond the current session for the User ID. Can you please check and help with this.
Have tried repackaging the project as well. Can you help please?
I have a Keynote Presentation coming up for the GARI International Conference. I am opening the conference and presenting a live demonstration.
Following up on the Long Term Memory issue. Despite manually resetting the User ID in the project and repackaging using SDK 3.6.8-beta.3, the issue persists for character 81b2b358-6cca-11f0-9210-42010a7be01f. Memory is still not being retained across sessions. Any further guidance would be appreciated.
The memory used in the llm context is visible to you in the mindview section.
Were you after a specific information that couldn’t be recalled?
From what we can see, this may be more of a recall/response quality issue than a memory availability issue. The memory snippets are provided to the LLM as additional context, but the model still decides how to use that context in its response. In some cases it may not surface a relevant memory, may prioritize the current conversation, or may miss a detail even though the memory was available in the prompt.
Thanks Ayush, really appreciate you taking the time to investigate this in detail.
That’s actually very insightful and aligns with some observations we’ve been exploring regarding relational continuity, memory prioritisation, and emotional persistence in empathic digital humans.
I suspect the issue may indeed be less about memory availability and more about how contextual and emotionally significant memories are being prioritised and interpreted at inference time.
I’ll send you a DM shortly with a bit more context from some proprietary research work I’ve been conducting around this area, as I think it may help explain some of the behavioural patterns we’re observing with the digital humans.
