Self-hosted large-document ingest fix: offload workflow step data past Rivet's 128 KiB KV limit, harden shutdown lifecycle. Fixes Windows build failure from rc.1.

Self-hosted supermemory can now use local or remote embedding models.

• Pluggable embeddings — local ONNX (default) or OpenAI / OpenAI-compatible / Google via SUPERMEMORY_EMBEDDING_*; first-boot picker + plan lock in embedding-plan.json
• Safe upgrades — populated stores without a lock assume legacy local · Xenova/bge-base-en-v1.5 · 768d; same-dimension model switches fail fast instead of mixing vectors

curl -fsSL https://supermemory.ai/install | bash

supermemory-server upgrade

Your local supermemory instance now turns documents into richer, more reliable memories — and gets you to searchable results faster.

• Smarter memory generation — documents produce fuller, more useful memories from the content you add
• Smoother ingestion — uploads move through the pipeline more reliably, with less time stuck in queue
• Better file handling — PDFs and spreadsheets are handled more predictably on upload
• Time-aware search — ask about memories from "last week" or a specific period
• Richer profiles — /v4/profile returns organized memory buckets
• Clearer search ranking — inferred memories stay available but rank below your core memories

curl -fsSL https://supermemory.ai/install | bash

supermemory-server upgrade

Telemetry is optional — set SUPERMEMORY_DISABLE_TELEMETRY=1 to turn it off completely.