Second Brain: Personal AI Operating System

Two Model Tiers, Voyage Embeddings, and Whisper

Claude Sonnet 4.6 handles all conversation, synthesis, and scheduled briefings. Claude Haiku 4.5 handles high-volume, low-stakes tasks: scoring exchanges for vault eligibility and classifying ingested documents. Voyage AI generates the embeddings behind semantic search, and OpenAI Whisper transcribes voice notes. Prompt caching on every Claude call trims repeat token cost. Running on Node.js 24 on a Hetzner CX23, this keeps quality high where it matters and total cost at approximately $15 to $25 per month at moderate daily use.

Smart Save Filter

Every exchange is scored 1, 2, or 3 by Haiku before any save decision is made. Score 1 (casual chatter, simple confirmations) is discarded. Score 2 or 3 triggers a vault write. A fail-safe saves on classifier error so no worthy exchange is ever lost. The vault stays high-signal automatically.

From Silent Extraction to an Intentional Task Store

The original design had Haiku silently extract tasks from every saved exchange. In practice, auto-extraction produced noisy, sometimes hallucinated tasks, so it was disabled and replaced with a deliberate SQLite-backed store: stable IDs, created, due, and completed dates, project, priority, and open/done/cancelled status, with the markdown task log kept as a projection. Every /done and /cancel passes a confirmation gate that shows the exact task and never bulk-completes, and cancelled stays distinct from done so task hygiene is never counted as completed work.

Multi-Turn Conversation Memory

The last 6 raw message turns are replayed into every Claude prompt, giving the model true short-term conversational context within a session. In-memory per-process, resets on restart. Vault context provides long-term persistence across sessions.

PARA Vault Migration

The original flat folder structure was migrated to a numbered PARA system (00-Inbox through 05-Templates) with all 14 bot.js write and read paths updated in the same coordinated deployment. A matched-count safety check ran before any old folder was deleted. 270 notes moved without loss.

ChatGPT History Import

A standalone importer reconstructed 201 conversations from a ChatGPT export, skipped 67 thin chats, wrote clean vault notes with a dry-run mode and timestamp collision guard, then removed itself and the export file from the server. Years of prior thinking became searchable context overnight.

axios over Anthropic SDK

The Anthropic SDK throws Premature close errors on the Hetzner host. All Claude API calls use direct axios HTTP with manual headers, a 90-second timeout, and 3 retry attempts. This is the confirmed working pattern for this server environment.

Scheduled Agents via node-cron

Three cron jobs run timezone-pinned to Asia/Dubai: morning briefing at 07:30, afternoon nudge at 15:00, and weekly review at 18:00 on Sundays. Each job is wrapped in try/catch with failure logging. ensureTaskFile() runs once at startup so the task layer is always ready before the first cron fires.

Chunked Semantic Search with Voyage AI

Notes are split into passages and embedded with Voyage AI into a vector index of ~2,300 curated chunked vectors stored in SQLite, with raw chat transcripts excluded at index time. Chunking lets long notes match on the specific passage that is relevant rather than an averaged whole-note vector. Retrieval ranks by cosine similarity above a similarity floor that filters weak matches, the index auto-reindexes whenever a note is saved, and stale vectors are pruned so search stays current with zero manual maintenance.

Multi-Format Ingestion Pipeline

PDF, PPTX, and DOCX uploads are parsed to text, then Claude classifies each document and files it into the correct PARA folder automatically. A vault note is written with structured analysis and the original is backed up to Google Drive, turning any dropped file into searchable, filed knowledge without manual sorting.

Branded Output Generation

The /deck, /report, /brief, and /onepager commands assemble vault content into finished documents rendered on a consistent charcoal, gold, and cream branded template. Generation pulls from indexed notes so outputs reflect what the system already knows rather than starting from a blank page.

Hosted MCP Server over Cloudflare Tunnel

A Model Context Protocol server exposes four tools, search_vault, get_note, list_assets, and save_note, over a permanent Cloudflare tunnel at vault.stevehaddad.io. Both the Claude app and Claude Code connect to it, giving every Claude surface authenticated read and write access to the same vault the Telegram bot uses.

API Spend Tracking

A shared spend.js module logs every Claude and Voyage AI call with token usage and cost, accumulating a running monthly total surfaced on demand via /spend. Spend is observable rather than a surprise at the end of the month, which is what keeps the whole system verifiably inside the $15 to $25 range.

Tool-Use Agent Loop

Conversation moved from static pre-injected context to a tool-use agent loop. Claude is given tools for vault search, web search, and spend lookup, and decides for itself what to pull and when, iterating until it has what it needs. Generators return structured outputs so decks, reports, and documents assemble from typed fields rather than parsed free text.

Voice Notes via Whisper

Telegram voice notes are downloaded, sent to OpenAI Whisper-1 for transcription, and fed into the exact same pipeline as a typed message, so a spoken question flows through semantic search, the agent loop, smart-save scoring, with no separate path.

Research and Competitive Intelligence

/research fans a topic out through Serper and synthesizes the results into a briefing. /competitive drives the Apify Website Content Crawler over a target company site and turns the crawl into a structured competitive readout. Either can chain into a branded deck or report in the same message.

Book Intelligence

A book PDF is parsed, split into sequential part notes, and given an index note that ties the parts together. Every part is embedded, so /book scopes retrieval and Q&A to one title while /listbooks enumerates the shelf. Long-form reading becomes queryable knowledge rather than a static file.

Two-Way Sync, Backups, and Version Control

rclone bisync mirrors the vault two ways with Google Drive so edits from any device converge. A private GitHub repo (stylo-dev-dxb/second-brain) holds the code and vault history, and a 2am cron pushes nightly backups to Drive with 7-day retention. No single machine or service is a point of failure.

Uptime Monitoring and WAF Hardening

The bot pings a healthchecks.io check on a heartbeat, so a silent crash raises an alert instead of going unnoticed. Cloudflare WAF rules sit in front of the MCP tunnel and site, hardening the public surface after the migration off Namecheap DNS.

Computed Health Baselines over WHOOP Data

/whoop and /whoop_deep pull recovery, HRV, resting heart rate, sleep, and strain from the WHOOP API into a dedicated SQLite time-series holding 90+ days of daily physiological records. 7, 28, and 90-day personal baselines, deviation from baseline, trend, and sleep debt are computed in code, so a recommendation reads "42% recovery, 14 points below your 28-day baseline" instead of a raw number with no context. Reasoning is grounded in computed signals, never a single unreferenced snapshot.

Retrieval Evaluation Harness

A 30-question evaluation set with expected-note matching measures retrieval quality objectively. The baseline scored 90%. Excluding chat-log noise lifted it to 97%, and enriching a key canonical note brought it to 100%. "Is retrieval good?" went from an assumption to a measured number, with regression protection on every future index change.

One Exclusion Predicate, Zero Index Drift

Raw chat transcripts made up 39% of the vector index and buried curated notes in retrieval, so 1,437 of those vectors were removed and the index-exclusion logic was unified into a single shared predicate, so the build and reindex paths can never drift apart. The WHOOP and task databases also joined the nightly backup rotation via the SQLite online-backup API.