Feature Request: Persistent Memories

[pleibers]

Usually some models tend to make the same mistake over and over again, also project specific information can be kept to avoid repeated searches etc.

A project memory with something like Serena does would be amazing, since mcps are not really the idea behind dirac.

So a simple project memory (much simpler than serena) would be very beneficial i think

[dirac-run]

This could potentially be a good idea and it is easy to implement in a naive manner (just supply a workspace scoped scratchpad). In the naive approach, the obvious issue is unbounded append only token bloat as agents will keep writing things that were important on some run but is junk for other runs. I am still thinking about how to solve this elegantly. I like the idea and happy to put it on the roadmap.

[dirac-run]

I spent quite some time on this. Most of the potential benefits of having persistent memories seem to be negated from the memories being too task specific but applied broadly out of context. Do you have any examples where the memory system adds and compounds value in the long run?