[GH-ISSUE #10895] Bypass vector-based retrieval of the most relevant memory, and instead populate the user's "Memory" content into the system prompt via variables. #16071

New Issue

2026-04-19T22:06:57-05:00

GiteaMirror commented

2026-04-19 22:06:57 -05:00

Originally created by @CookSleep on GitHub (Feb 27, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/10895

Is your feature request related to a problem? Please describe.
I've noticed that Open WebUI's "Memory" feature works by querying for the most relevant memory item on each request. However, this approach is quite limiting. In my personal experience, injecting all user memories directly into the system prompt is much more effective.

Describe the solution you'd like
I'd like to see support for an option to bypass vector-based retrieval of the most relevant memory. Instead, all user "Memory" content should be populated into the system prompt via variables.

Describe alternatives you've considered
I've tried a couple of community-made memory tools. They're cool, and I've taken the liberty of modifying and extending the latest version of one of them using Windsurf (the community hasn't fully developed this better new version). If possible, I'll release it to the community soon.
However, this tool can only send all memories to the model when the model remembers to call the "return all memories" function. This isn't conducive to natural conversation.

Additional context
Memory is a great tool, and I hope you'll have time to add formal, complete support for it, similar to web search queries!

Originally created by @CookSleep on GitHub (Feb 27, 2025). Original GitHub issue: https://github.com/open-webui/open-webui/issues/10895 **Is your feature request related to a problem? Please describe.** I've noticed that Open WebUI's "Memory" feature works by querying for the most relevant memory item on each request. However, this approach is quite limiting. In my personal experience, injecting all user memories directly into the system prompt is much more effective. **Describe the solution you'd like** I'd like to see support for an option to bypass vector-based retrieval of the most relevant memory. Instead, all user "Memory" content should be populated into the system prompt via variables. **Describe alternatives you've considered** I've tried a couple of community-made memory tools. They're cool, and I've taken the liberty of modifying and extending the latest version of one of them using Windsurf (the community hasn't fully developed this better new version). If possible, I'll release it to the community soon. However, this tool can only send all memories to the model when the model remembers to call the "return all memories" function. This isn't conducive to natural conversation. **Additional context** Memory is a great tool, and I hope you'll have time to add formal, complete support for it, similar to web search queries!

GiteaMirror closed this issue

2026-04-19 22:06:57 -05:00

GiteaMirror commented

2026-04-19 22:06:58 -05:00

@alghostzx commented on GitHub (Feb 27, 2025):

agreed, several times there's many relevant memories, but the current system seems focused on only ever feeding a specific one

@alghostzx commented on GitHub (Feb 27, 2025): agreed, several times there's many relevant memories, but the current system seems focused on only ever feeding a specific one