It would help if there was a node for:
chat Memory node for Mem0
My use case:
Mem0 offers additional functionality compared to the existing options
Did you make any progress on this idea? I think itâs a very good one.
I built a memory system myself using other n8n nodes (info extractor, Postgres, memory manager) instead of mem0. It extracts information in a conversation based on categories stores and updates it in Postgres. Some info I retrieve before inference and insert to the system message, some the agent can retrieve if necessary.
Hey Paul, would you be willing to share a bit more about how you set this up? I started going down a similar path by building a really simple system to write & recall 1-sentence memories (inspired by youtube videos) and now Iâm also storing things like last-thread interaction on slack, but I want to expand into what youâre talking about which is categorical memories.
Are you utilizing a vector store under the hood with chunking for this?
Hey envisean,
It is not a very sophisticated approach and not much more than the 1-sentence memories you mentioned. No vector store, but plain text in a PostgresDB.
Here is how it works:
After response was send, memories are extracted with the info extractor and categorized. Make sure you have really specific instructions for the categories. The memory updater is just a tools agent instructed to check the memories of this user and add/update if necessary. For each category there is something like this in the AGent
1. Update the "category_name _here" with this info: "{{ $json.output.category_name_here }}" according to the following rules:
- Do not add redundant information
- Only remove information from the "personal details" if the new information is updating this info and contradicting the existing info.
- Information must be written in third person style such as "has a wife, "likes strawberries", "is tired"
You can retrieve the memories and insert into the system prompt or hand the agent access to postgres as a tool and let it check for relevant memories in a relevant category.
Be aware that the updater agent could delete/overwrite memories by mistake and there is a possibility that a user would be able inject malicious instructions through the memory text somehow. Its enough for my usecases though.
Hope it helps/provides some inspiration.
1 Like
Yeah, this is actually super helpful to see your approach.
Youâve got me questioning the amount of tools I gave to my single assistant agent. Probably will try something similar so I can get an immediate response back to the user while it finishes up the the memory writing. Thanks!
1 Like
I just came across mem0 as well. Keen to see if anyone has implemented it on n8n yet
we are waiting mem0 connector
You can now use mcp to run Mem0 in your workflow:
You will need to modify main.py to add the user_id parameter because right now the main.py file from the github repo defaults user_id to âuserâ, kinda uselessâŚ
Any updates on this?
Please click the vote button above this post on the left side, this will help developers prioritize mem0 integration.