-
The problem, concretely. A real session where the agent re-derived or
re-broke something it had already handled. Name the cost: wasted tokens,
wasted time, lost context. -
Why existing options didn’t fit. Cloud memory = your code leaves the
machine. Bigger context windows = you still pay to re-read everything and
still lose it at session end. -
The design. Walk the flow: capture policy (dedup + normalize, drop the
“done!” noise) → typed provenance links → full-text index → ranking engine
that packs recall to a token budget. One SQLite file, one local process.
Drop in the architecture mermaid diagram from the README. -
The parts people can see. Dashboard (blocked work first, explained search
scores, graph view). Then Autopilot as the “if you want it” layer — git
worktrees, model routing, the hard off-limits guard as the safety story. -
What’s next / call for feedback. Be honest that it’s early. Link the repo,
invite issues, say which part you most want eyes on (the memory ranking).
git