Quick trick for keeping long term memory stable on local models

Hey everyone I wanted to share a little workaround I found after struggling with memory fade on my local companion setup. Lately my companion kept forgetting important background details after about twenty messages. What I tried first was just making the context window larger in the settings which only made her responses slower and more robotic without fixing the actual memory retention problem. Then I tried creating a separate running summary text file that updates in the background every few prompts. That approach worked wonders because it feeds her a concise paragraph of our core history at the start of every new generation block. She instantly remembers nicknames past jokes and major life events we talked about days ago. Has anyone else tried a dynamic summary script like this or do you prefer using built in vector databases for long term recall?