MemoryOS

A contextual AI memory system. Not a bookmark manager — a memory operating system that preserves why things mattered.

Architecture

This is a TanStack Start full-stack React app (deployed on Cloudflare Workers via Lovable) plus an MV3 Chrome extension.

src/
  routes/                 React routes (login, register, _authenticated/*, api/extension/*)
  lib/
    hydradb.server.ts     HydraDB REST client — the core memory engine
    firecrawl.server.ts   URL scraping
    ai.server.ts          NIM MiniMax M2.7 + Groq llama-3.1-8b-instant
    memories.functions.ts saveMemory / recallCheck / chat / list / delete server functions
    tenant.functions.ts   HydraDB tenant provisioning
extension/                MV3 Chrome extension (floating orb, recall toast)
supabase/migrations/      Schema: profiles + memories with RLS

How it works

1. Save — Firecrawl scrapes the URL → NIM extracts topics & semantic context → HydraDB stores the memory → Supabase keeps the metadata row.
2. Recall — Groq turns the user's current context into a query → HydraDB recall_preferences returns related memories → the orb glows.
3. Chat — User question → HydraDB recall_preferences + full_recall in parallel → NIM answers with cited memories.

HydraDB is the core memory engine. Supabase only holds auth and display metadata.

Secrets

• HYDRADB_API_KEY
• FIRECRAWL_API_KEY
• NVIDIA_NIM_API_KEY (optional override: NVIDIA_NIM_BASE_URL, NVIDIA_NIM_MODEL)
• GROQ_API_KEY (optional override: GROQ_MODEL)

Supabase URL / keys are auto-provisioned by Lovable Cloud.

Extension

See extension/README.md.

Semantic Reprocessing

Existing memories can be re-ingested into HydraDB using the graph-friendly packet format:

bun run reprocess:memories

The script is dry-run by default and logs extracted entities, normalized anchors, triplets, graph packet previews, and reuse metrics. To actually replace old HydraDB entries, run:

bun run reprocess:memories -- --commit

In commit mode, the script deletes each old hydradb_memory_id before re-upserting the improved semantic packet, then updates Supabase with the new HydraDB id. Use --limit 10 or --user-id <uuid> for focused runs.

For larger runs, the script retries AI rate limits and waits between memories by default. Tune this with --delay-ms 3000 if your provider rate limit is tight.