Stashov — Personal AI Memory

A personal AI that remembers what you want to buy and what you know. It saves links, finds them on demand, watches prices, and nudges you when it's time your data, on open infra you own.

Comment

Stashov — Personal AI Memory & Impulse Gate

Inspiration

I've spent years accumulating digital artifacts: 1,346 browser bookmarks across Chrome, Brave, and Edge; camera-roll screenshots of products, recipes, and code snippets; saved links from TikTok, YouTube, Reddit, and X. The problem wasn't saving — it was retrieval and context.

Standard bookmark folders and wishlist apps fail in two ways:

  1. Retrieval & Link Rot: Saved content drowns in unorganized folders. Links rot. Screenshots sit in camera rolls with no semantic connection to what they represent.
  2. Zero Personal Context: A wishlist doesn't know I already own a Black Quilted Winter Parka. It doesn't know my evening routine uses 0.1% Adapalene and that a viral glycolic cleanser would conflict. It doesn't know I bookmarked four fragrance discounters carrying the same cologne for $87 less.

Stashov was born from a simple question: What if my digital memory actually understood me — my inventory, my constraints, my budget, my routines — and used that context to gate every impulse before it became a purchase?

What it does

Stashov is a local-first Personal AI that unifies your saved bookmarks, wishlist, living routines, and personal constraints into one connected brain.

Capability: Browser Bookmark Vault Description: Imports 1,346 bookmarks from Chrome/Brave/Edge, deduplicates via canonical URL, classifies into 14 categories (Programming, AI, Menswear, Fragrances, Skincare, Watches, Soccer, Sourcing, etc.) ──────────────────────────────────────── Capability: Reference Library Description: 1,017 non-shopping links organized into 10 sections (Education, Programming, AI, Career, Business, Design, Cyber, Utilities, Gaming, Sourcing) with live health status (858 live, 110 blocked, 61 removed, 9 restored after review) ──────────────────────────────────────── Capability: Impulse Gate Engine Description: 14-day cooling-off state machine (cooling → ready → bought/skipped/snoozed), monthly ledger with editable caps, duplicate overlap check against owned inventory, discounter/dupe cross-reference, price history with sparklines and target alerts ──────────────────────────────────────── Capability: Living Protocols Description: Paste-to-routine builder (TikTok/YouTube caption → AM/PM steps + guardrails), product linking with cached photos/verified prices, conflict detection (blocks AHA/BHA in PM alongside 0.1% Adapalene), daily check-off log with streak tracking ──────────────────────────────────────── Capability: Capture from Anywhere Description: Desktop bookmarklet (#captureDialog) reads JSON-LD/OpenGraph client-side, paste-to-gate, mobile PWA with Web Share Target at /share-target ──────────────────────────────────────── Capability: Nebius Serverless Worker Description: Deal/discounter polling, link-rot verification, cooling-off unlocks, weekly digest compilation ──────────────────────────────────────── Capability: Telemetry Console Description: Live Nebius Token Factory routing latency, token usage, cost per call

How we built it

Architecture

┌─────────────────────────────────────────────────────────────────┐ │ STASHOV VAULT │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │ │Bookmarks │ │ Wishlist │ │Protocols │ │ Profile │ │ │ │ (1,346) │ │ (13 items)│ │ (5) │ │ (sizes, │ │ │ │ 14 cats │ │ + gate │ │ + daily │ │ constraints,│ │ │ └────┬─────┘ └────┬─────┘ └────┬─────┘ │ inventory)│ │ │ │ │ │ └────┬──────┘ │ │ └─────────────┼─────────────┼──────────────┘ │ │ ▼ ▼ │ │ ┌──────────────────────────────┐ │ │ │ IMPULSE GATE ENGINE │ │ │ │ • Duplicate overlap check │ │ │ │ • 14-day cooling-off FSM │ │ │ │ • Monthly ledger + caps │ │ │ │ • Discounter/dupe cross-ref │ │ │ │ • Price history + targets │ │ │ └──────────────┬───────────────┘ │ │ ▼ │ │ ┌──────────────────────────────┐ │ │ │ NEBIUS TOKEN FACTORY │ │ │ │ ┌────────────────────────┐ │ │ │ │ │ nvidia/Nemotron-3-Ultra│──┼──► 1M-context synthesis│ │ │ │ nvidia/Nemotron-3-Super │──┼──► Agentic synthesis │ │ │ │ nvidia/Nemotron-3-Nano │──┼──► Fast extraction │ │ │ │ nvidia/Nemotron-3.5-Lightning│ MTP fast decode │ │ │ │ Qwen/Qwen3-Embedding-8B │──┼──► Semantic memory │ │ │ │ zai-org/GLM-5.3-Flash │──┼──► Vision OCR │ │ │ └────────────────────────┘ │ │ │ └──────────────────────────────┘ │ └─────────────────────────────────────────────────────────────────┘

Core Stack

Layer: Runtime Technology: Node.js 18+ (raw http module — no framework) ──────────────────────────────────────── Layer: AI Platform Technology: Nebius Token Factory (OpenAI-compatible REST) ──────────────────────────────────────── Layer: Models Technology: Nemotron-3 Ultra/Super/Nano/Lightning, Qwen3-Embedding-8B, GLM-5.3-Flash ──────────────────────────────────────── Layer: Storage Technology: JSON vault (2.5MB) + atomic writes (temp + fsync + rename + .bak + 20 snapshots); Vercel Blob in production ──────────────────────────────────────── Layer: Background Technology: Nebius Serverless Jobs (periodic workers, no infra) ──────────────────────────────────────── Layer: Frontend Technology: Vanilla JS SPA (8 tabs), CSS design system, PWA manifest + Web Share Target ──────────────────────────────────────── Layer: Security Technology: Loopback binding, HMAC-SHA256 sessions, CSP/COOP/CORP/HSTS, SSRF-safe fetch, rate limits ──────────────────────────────────────── Layer: Deploy Technology: Vercel (primary), Render (alt), Nebius Serverless Endpoints (native)

Key Modules

File Responsibility
scripts/sync-bookmarks.js Chrome/Brave/Edge importer + 14-category classifier
scripts/audit-library.js Link health check + description extraction
lib/gate.js 14-day FSM, ledger, budget caps, buy/skip/snooze/reopen/void
lib/promotion.js Pre-gate review, JSON-LD/OpenGraph capture, price provenance
lib/routine.js Paste-to-routine parser, conflict guardrails
lib/safe-fetch.js SSRF-safe HTTP (DNS pinning, private IP blocking, size/time limits)
lib/state.js Atomic vault writes, backups, hosted storage abstraction
jobs/nebius-serverless-worker.js Background deal polling, link-rot, unlocks, digests

Challenges we ran into

  1. SSRF-Safe Fetching Without Breaking Legitimate Use Product pages must be fetched server-side for metadata, but that opens SSRF vectors. Solution: lib/safe-fetch.js — every redirect revalidated, DNS resolved and pinned, private/loopback/link-local/metadata addresses blocked, credentials and non-standard ports rejected, response size/time limits enforced.

  2. Price Extraction Confidence Without Guessing Stores return 403 to server requests; JSON-LD absent; visible prices are ranges, "from" prices, installment copy. Solution: Bookmarklet reads DOM client-side, prefers sale price over compare-at, follows selected variants, explicitly marks low confidence in review/gate card/saved note — never fabricating a $95 placeholder.

  3. Atomic Writes at Scale on Serverless Vault is 2.5MB JSON. Vercel Blob + serverless = no local fsync. Solution: lib/state.js abstracts storage — local uses temp-file + fsync + rename + .bak + timestamped snapshots (last 20); hosted uses Vercel Blob with pre-import snapshots. Telemetry no longer rewrites vault on every search.

  4. Model Routing Without Hardcoding Nebius Token Factory exposes /api/public/models_info with capabilities (function_calling, context_and_rag, responses_api, MTP). Solution: Data-driven capability map in server.js selects cheapest model satisfying task requirements.

  5. Duplicate Detection That Doesn't Annoy Title similarity (Jaccard on filtered tokens) + same-domain heuristic + active-wishlist filter. Threshold: 0.65 similarity OR exact title match. Surfaces top 4 matches in review dialog before the 14-day clock starts.

Accomplishments that we're proud of

Achievement: 1M-context synthesis works Impact: Nemotron-3-Ultra processes 1,346 bookmarks + Reddit threads + ingredient sheets + inventory + budget + routines in one pass — genuine multi-document reasoning ──────────────────────────────────────── Achievement: 10× cost reduction via tiered routing Impact: Ultra (deep) → Super (agentic) → Nano (fast) → Lightning (MTP) → Qwen (embeddings) → GLM (vision) — right model per task ──────────────────────────────────────── Achievement: Zero-config local bookmark import Impact: npm run sync-bookmarks scans Chrome/Brave/Edge stores automatically — no manual export/upload ──────────────────────────────────────── Achievement: Bookmarklet that works on 403 pages Impact: Client-side DOM capture falls back when server requests fail, with confidence labeling ──────────────────────────────────────── Achievement: Conflict detector that knows your routine Impact: Blocks AHA/BHA in PM window alongside 0.1% Adapalene — personal context as executable guardrails ──────────────────────────────────────── Achievement: Audit trail on every mutation Impact: Ledger entries, price observations, gate transitions, backup snapshots — everything auditable ──────────────────────────────────────── Achievement: Live demo on Vercel with Blob Impact: https://stashov.com — /api/health 200, /api/vault 401 signed out, unlocks with token ──────────────────────────────────────── Achievement: 13-test suite covering security + logic Impact: Gate lifecycle, product extraction, promotion review, capture DOM, search relevance, SSRF guards, routine daily log, profile, state durability, unlock rate-limits

What we learned

Lesson: 1M-token context isn't a gimmick Evidence: Enables synthesis across entire personal corpus — the differentiator for personal AI ──────────────────────────────────────── Lesson: Tiered routing beats single-model Evidence: 10× latency/cost improvement; capability-driven selection via Nebius model info endpoint ──────────────────────────────────────── Lesson: Local-first + serverless = privacy + scale Evidence: Loopback binding, HMAC sessions, atomic writes; Nebius Jobs for background without infra ──────────────────────────────────────── Lesson: Impulse gate is a state machine, not a prompt Evidence: Deterministic transitions from timestamps — no LLM hallucination of time ──────────────────────────────────────── Lesson: Personal context must be editable, not hardcoded Evidence: Profile editor validates, protects budget/cooling fields, preserves IDs, writes atomically ──────────────────────────────────────── Lesson: Confidence labeling > hallucinated data Evidence: Low-confidence price marks propagate through review → gate card → saved note — user decides ──────────────────────────────────────── Lesson: Serverless Jobs fit periodic personal workloads Evidence: Deal polling, link-rot, unlocks, digests — no always-on server needed

What's next for Stashov — Personal AI Memory

Near-term (Hackathon → v2.1)

  • SAM 3.1 integration: Zero-shot segmentation on screenshot ingestion → auto-extract product regions, prices, specs
  • Muse Image: Generate contextual illustrations from retrieved memories (e.g., "show me that fragrance layering combo")
  • Muse Code: On-demand code synthesis from past snippets in Programming cluster
  • 3-minute demo video for submission

Medium-term (v2.2–v3.0)

  • Cross-device sync via encrypted conflict-free replicated data type (CRDT)
  • Voice capture (Whisper on Nebius) → ingest while walking/driving
  • Proactive agent: "You've saved 3 colognes from FragranceNet — Olfactory Factory has the same one for $87 less. Want me to gate it?"
  • Fine-tuned Nemotron Impulse Gate distilled from Ultra prompts → smaller, faster, specialized
  • WebSocket streaming on Token Factory for perceived latency on 1M-context calls

Long-term vision Stashov becomes the personal context layer that any AI assistant (local or cloud) can query — your inventory, constraints, routines, budget, and memory — with the impulse gate as the universal "pause and think" mechanism before any action commits resources.

Built With

Category: Languages Technologies: JavaScript (Node.js 18+), HTML5, CSS3 ──────────────────────────────────────── Category: AI/ML Platform Technologies: Nebius Token Factory, NVIDIA Nemotron 3 ──────────────────────────────────────── Category: Models Technologies: nvidia/Nemotron-3-Ultra-550b-a55b, nvidia/Nemotron-3-Super-120b-a12b, nvidia/Nemotron-3-Nano-30B-A3B, nvidia/Nemotron-3.5-Lightning, Qwen/Qwen3-Embedding-8B, zai-org/GLM-5.3-Flash ──────────────────────────────────────── Category: Cloud/Hosting Technologies: Vercel, Render, Nebius Serverless Endpoints ──────────────────────────────────────── Category: Storage Technologies: Vercel Blob, JSON + atomic writes, Nebius Serverless Jobs ──────────────────────────────────────── Category: APIs Technologies: Nebius Token Factory REST, Chrome/Brave/Edge bookmark APIs, Web Share Target ──────────────────────────────────────── Category: Security Technologies: HMAC-SHA256, CSP, COOP/CORP, HSTS, SSRF-safe fetch, rate limits ──────────────────────────────────────── Category: Testing Technologies: Node.js native --test (13 test files)

Try It Out

Link Description
🌐 Live Demo https://stashov.com
📁 GitHub Repo https://github.com/Arafat87/stashov-personal-ai
🏥 Health Check https://stashov.com/api/health

HOW TO ACCESS: STASHOV_ACCESS_TOKEN: 9c702e171aa063819a576de7b022992c96b0de0b8a2844f5cae7a37bfc24f3c3

bash git clone https://github.com/Arafat87/stashov-personal-ai.git cd stashov-personal-ai npm ci cp .env.example .env # Add NEBIUS_API_KEY npm run sync-bookmarks npm start # http://localhost:3847

Mathematical Appendix: Gate State Machine

The impulse gate operates as a deterministic finite automaton with timestamp-derived transitions:

$$ \begin{aligned} \text{States} &= {\texttt{cooling}, \texttt{ready}, \texttt{bought}, \texttt{skipped}, \texttt{snoozed}} \ \text{Transitions} &: \ \texttt{cooling} &\xrightarrow{t \geq \texttt{unlock_at}} \texttt{ready} \ \texttt{ready} &\xrightarrow{\text{buy}} \texttt{bought} \quad (\text{ledger entry written}) \ \texttt{ready} &\xrightarrow{\text{skip}} \texttt{skipped} \ \texttt{ready} &\xrightarrow{\text{snooze}(d)} \texttt{snoozed} \quad (\texttt{unlock_at} \leftarrow \texttt{unlock_at} + d) \ \texttt{snoozed} &\xrightarrow{t \geq \texttt{unlock_at}} \texttt{ready} \ \texttt{bought} &\xrightarrow{\text{reopen}} \texttt{ready} \quad (\text{ledger entry voided}) \ \texttt{skipped} &\xrightarrow{\text{reopen}} \texttt{ready} \ \end{aligned} $$

Built With

Share this project:

Updates

Submission history