i dont understand how this helps? i can have a folder with all md files, tell any agent "refer this folder for past context, learnings, and update/ improve this as we work, this is your memory". you are holding a sqlite for index search, but maintaining as the memory grows would be the hard part. Some sort of pruning/ dreaming is required always, beyond certain time.
At small size folder .md works well, we measured it. The difference is when the memory grows and you have multiple agents, the agent gets 5k relevant tokens instead of reading everything (98.4% vs 60% on LongMemEval) the memory works in most agents, and corrections replace old facts while keeping history. You're right about pruning, we have a feature like dreaming that works on that.
i dont understand how this helps? i can have a folder with all md files, tell any agent "refer this folder for past context, learnings, and update/ improve this as we work, this is your memory". you are holding a sqlite for index search, but maintaining as the memory grows would be the hard part. Some sort of pruning/ dreaming is required always, beyond certain time.
At small size folder .md works well, we measured it. The difference is when the memory grows and you have multiple agents, the agent gets 5k relevant tokens instead of reading everything (98.4% vs 60% on LongMemEval) the memory works in most agents, and corrections replace old facts while keeping history. You're right about pruning, we have a feature like dreaming that works on that.