docs: foreground memory + AutoResearch; reframe as a model you work WITH

[JeffBrines]

What

Two README fixes from feedback: memory and Karpathy's AutoResearch were present but buried, and the opening framed the model as a static deliverable.

Changes

• A model you work with, alongside the agent — new opening paragraph: ask "what if," challenge an assumption, flag something off, and the agent edits the model, re-runs, and explains. Pair-modeling, not waiting on a report.
• Memory, foregrounded — pulled up to the opening ("It also remembers…") and made a named pillar in the Mission ("A memory that compounds"). It was previously only a sub-clause + a roadmap row.
• AutoResearch, foregrounded — the self-improving-loop lineage is now stated up top (AutoResearch tunes against validation loss; openfpa against reconciliation error on your books), and the deeper "How it compounds" reference trimmed so it reinforces rather than repeats.

Tone stays humble and emoji-free. Docs only; suite unaffected.

🤖 Generated with Claude Code