v1.5.0 — Memory Architecture Overhaul

What's new in v1.5.0

Memory now actually works

• Direct fact extraction — facts you share in conversation are extracted automatically and stored directly to memory (no longer dependent on the background loop to catch them)
• Separated pipelines — user facts (source: conversation) vs AI thoughts (source: background_loop) are now distinct
• Follow-up questions — the background loop now generates follow-up questions about things you mentioned, stored separately from thoughts
• Re-engagement — when you return after 30+ minutes, Vespera surfaces a follow-up question naturally at the end of its response

Fixes from external tester report

• Periodic pruning crashed silently on startup (missing Path import)
• Shutdown orphaned the main process — now terminates cleanly
• DuckDuckGo search returned zero results (package renamed to ddgs) — fixed
• Background prompt was filtering OUT user facts — completely rewritten
• README listed shipped UI tabs as 'planned but not built'

Other fixes

• Cleanup crew no longer deletes user preference facts (was only keeping 'technical insights')
• MAX_THOUGHT_LENGTH raised 300 → 800
• Footer memory counts were showing dashes (wrong API level)
• Models page showed NaN MB for model sizes
• Background loop uses random memory sampling — scales to large memory stores without overflowing local model context
• 1399 junk memories from old broken prompt wiped on fresh install