Morning. I like the concept of your graph engine for opencode's memory. I independently came up with the graph idea for a harness I'm writing for role play, and I've had good results with it. The concept is that it commits as much knowledge as it can to graphs and then limits the conversation history it has in memory to the last 8 turns. It does a great job of remembering things that were mentioned even once or twice dozens of turns ago as the graph builds out over time. This includes references to the total conversation history by line number (like on line 75, it was mentioned that a particular character went to a place) so it has total recall of that conversation through simple graph recall and line retrieval.
It keeps my context windows super low. Normally over a role play session that's 100 posts long you would lose the beginning of the conversation in SillyTavern, and that would creep onward as you kept posting. I think the largest context window I've ever had after a few hundred turns is just over 10k tokens, and about 5-6k is persistent token usage from instructions and baseline knowledge.
The point is, I wonder if this could basically do the same thing. Limit the context window so it doesn't have to keep compacting, but include relevant file information as the LLM is making edits for things using a graph recall and file load system. Graph traversal is entirely local, but graph updating is partially LLM driven, so this means a slight token usage increase overall. But given that everything sent in context is relevant to the current operation and there is both a recent conversation history between the user and the LLM, it should improve the harness' ability to stay on task without becoming dumb for a significantly longer amount of time.
Sorry that I couldn't mark this as a feature request, but apparently I lack the option to do so.
Morning. I like the concept of your graph engine for opencode's memory. I independently came up with the graph idea for a harness I'm writing for role play, and I've had good results with it. The concept is that it commits as much knowledge as it can to graphs and then limits the conversation history it has in memory to the last 8 turns. It does a great job of remembering things that were mentioned even once or twice dozens of turns ago as the graph builds out over time. This includes references to the total conversation history by line number (like on line 75, it was mentioned that a particular character went to a place) so it has total recall of that conversation through simple graph recall and line retrieval.
It keeps my context windows super low. Normally over a role play session that's 100 posts long you would lose the beginning of the conversation in SillyTavern, and that would creep onward as you kept posting. I think the largest context window I've ever had after a few hundred turns is just over 10k tokens, and about 5-6k is persistent token usage from instructions and baseline knowledge.
The point is, I wonder if this could basically do the same thing. Limit the context window so it doesn't have to keep compacting, but include relevant file information as the LLM is making edits for things using a graph recall and file load system. Graph traversal is entirely local, but graph updating is partially LLM driven, so this means a slight token usage increase overall. But given that everything sent in context is relevant to the current operation and there is both a recent conversation history between the user and the LLM, it should improve the harness' ability to stay on task without becoming dumb for a significantly longer amount of time.
Sorry that I couldn't mark this as a feature request, but apparently I lack the option to do so.