TinkerUp User Guide

Overview

TinkerUp is a team-memory platform. Two modules work standalone or together:

Quick Start

1. Run only TinkerMeet (record + MoM, no memory)
   Start tinkermeet/backend + tinkermeet/frontend. Record audio in the browser, download or post MoM. No Slack or DB needed beyond what TinkerMeet ships.
2. Run only TinkerMem (Slack memory, no meetings)
   Set up Slack app + Google Drive service account + Neon Postgres. Run cd tinkermem && docker compose up -d --build. Bot listens to allowed channels.
3. Run combined (full team memory)
   Both modules + dashboard. TinkerMeet pushes MoM to TinkerMem; queries see Slack + meetings in one memory graph.

Module Choice

Pick by need, not by default. Each module is independent.

Slack Integration (TinkerMem)

TinkerMem connects to Slack via Socket Mode — outbound-only WebSocket. No public URL or inbound rules needed — runs fine in a private VPC.

What gets ingested

• Every message in allowed channels (SLACK_CHANNEL_IDS env)
• Google Drive links (auto-extracted via service account)
• Slack file attachments (docx/pdf/xlsx/etc — needs files:read scope)
• Emoji reactions on bot answers — used for feedback (👍/👎)

Setup steps

1. Create a Slack app
   api.slack.com/apps → New App → From scratch → enable Socket Mode → generate app-level token.
2. Add Bot Token Scopes
   app_mentions:read, channels:history, channels:read, chat:write, commands, files:read, users:read, reactions:read
3. Register slash commands
   /ask, /backfill, /recall, /correct, /usage (placeholder URL is fine — Socket Mode skips it).
4. Install + invite bot
   Install app to workspace, copy Bot Token (xoxb-) into .env. In channel: /invite @YourBot.
5. Backfill existing history
   In the channel, run /backfill to pull entire history into memory.

Meeting Recording (TinkerMeet)

Record + transcribe + generate MoM in the browser.

1. Start the meet UI
   cd tinkermeet/frontend && npm run dev. Open in browser, grant microphone permission.
2. Record audio
   Click Record. Audio streams over WebSocket to backend → OpenAI gpt-4o-transcribe.
3. Get auto-MoM
   When you stop recording, the bot sends transcript to Bedrock Haiku for MoM generation (decisions + action items + open questions).
4. Optional: post to Slack
   Set SLACK_MOM_CHANNEL in tinkermeet/.env. MoM auto-posts after generation.

Projects & Scoping

Projects are optional groupings. By default everything is standalone (independent). Assigning a Slack channel to a project gives you scoped queries — no cross-project leakage.

Rules

• Channel → project mapping is mutable. Reassigning affects future messages only.
• Each raw_message records the project at INSERT time — immutable after.
• Query with project_id → strict WHERE project_id = $X across all retrieval paths.
• NULL project_id = standalone. Standalone messages do not appear in project-scoped queries.

Manage projects

Go to /projects on the dashboard: create projects, assign Slack channels, see channel and message counts per project.

Memory Graph

Three tiers, each with confidence + temporal decay.

Why temporal decay matters

Each node has a confidence_scorethat decreases with age + conflicting newer context. Below threshold → filtered out of retrieval. Notion/Confluence don't have this — pages rot silently. Otter/Fireflies don't either — transcripts sit flat in folders.

File Extraction

Two ingestion paths, both end in file_chunks table for vector RAG.

Slack file uploads

Files dragged into a channel are downloaded via files.info + url_private_download (needs files:read scope). Supports docx, pdf, xlsx, html, md, jsx, and more.

Google Drive links

Drive URLs in chat are detected via regex, then content is exported via the Google Drive API. Requires sharing the file (or parent folder) with the service account email— the SA doesn't inherit creator permissions.

RAG flow

Each chunk is prepended with metadata (uploader + file name) before embedding so queries like "what did Nita send?" match by author.

Slash Commands & Mention

Web Chat

Same query engine, different UI. Open /chat on the dashboard. Optionally enter a Channel ID to auto-resolve project scope.

Cost Tracker

Open /costs for centralized LLM usage:

• Pie chart — cost contribution per model
• Bar chart — input vs output tokens per model
• Breakdown table — provider, model, calls, tokens, cost
• Period selector — today, 7d, 30d, all

Tinkermem logs internally; TinkerMeet POSTs to tinkermem/api/usage after each OpenAI/Bedrock call.

Quick Reference

Ports

Routes

Common gotchas