The 'store everything vs. synthesize' axis is the interesting design question here. MemPalace bets that retrieval quality is the problem — keep everything verbatim, make it findable. That's a legitimate approach.
We went the other direction with Lodestar (launching tomorrow). The bet is that you don't need the transcript but you do need the decision layer. What was decided, why, what was rejected. Synthesized by an LLM at session end into a plain markdown file that commits to the repo.
The other structural difference: the file lives in the codebase, not on the machine. git clone on a new laptop and the context is already there. Different tradeoff than a local vector DB.
Curious whether MemPalace users find the verbatim storage approach holds up as session count grows — the retrieval quality question gets harder with more history.