Open custom memory loading for LLM nodes #6

New Issue

2026-02-16T05:18:51-05:00

yindo commented

2026-02-16 05:18:51 -05:00

Originally created by @WenXIN-AI on GitHub (May 8, 2025).

Although chatflow supports session variable storage memory, the effect in actual application is still not as good as the built-in message adding method of LLM node.

Directly loading external memory as context into the prompt word of the large model is not a data format that the large model itself can handle well. The best way is to open the memory loading node.

I found that the memory processing of dify stayed in a simple context window for a long time and could not be loaded dynamically, which limited the search and other optimization processing of short-term memory. At the same time, if multiple models are tested and replied together, how to share the memory between them?
I hope to load the conversation context in the form of messages, because this is the data organization format that the large model has trained. If it is added directly to the prompt, it often has poor results and repeats the previous answer output.

I tried to modify the LLM node, but I am not familiar with the front-end language, and adding a tool is also a very troublesome thing. If the official updates some new features, everything has to be done again, so I hope to update a custom memory LLM node so that the LLM node can truly become a model reasoning service provider, rather than limiting its flexibility because of memory.

Originally created by @WenXIN-AI on GitHub (May 8, 2025). Although chatflow supports session variable storage memory, the effect in actual application is still not as good as the built-in message adding method of LLM node. Directly loading external memory as context into the prompt word of the large model is not a data format that the large model itself can handle well. The best way is to open the memory loading node. I found that the memory processing of dify stayed in a simple context window for a long time and could not be loaded dynamically, which limited the search and other optimization processing of short-term memory. At the same time, if multiple models are tested and replied together, how to share the memory between them? I hope to load the conversation context in the form of messages, because this is the data organization format that the large model has trained. If it is added directly to the prompt, it often has poor results and repeats the previous answer output. I tried to modify the LLM node, but I am not familiar with the front-end language, and adding a tool is also a very troublesome thing. If the official updates some new features, everything has to be done again, so I hope to update a custom memory LLM node so that the LLM node can truly become a model reasoning service provider, rather than limiting its flexibility because of memory.

yindo closed this issue

2026-02-16 05:18:51 -05:00

yindo referenced this issue

2026-02-16 05:19:57 -05:00

[PR #6] [MERGED] Feat: update link convertor #308

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: langgenius/dify-docs#6