Obsidian FIXED My OpenClaw Memory Loss! — Superbash Learn

Superbash Learn

Click to load the YouTube player

Key takeaways

Run a daily cron at your local time (Ron uses 7 a.m. Hong Kong, not UTC) and dump each day's full chat into a markdown file in your Obsidian vault.
Back the vault up to GitHub so the journal survives — this is what gives the agent a readable, auditable history instead of opaque vector embeddings.
Install Smart Connections (Brian Petro, 836k downloads) for local, private embeddings, and QMD as MD to lock the agent into a clean markdown schema.
Use Discord's per-topic channels instead of WhatsApp/Telegram — one context per workflow stops the "muddied" single-chat problem Ron hit in two days.
Split your setup into two jobs: a 2-minute Brave Search daily brief (X + Finviz + TechCrunch + crypto sites) and a separate deep-research flow that fires parallel sub-agents.
Feed the agent your own trade log — entries, order flow, volume, delta — and ask it to find recurring variables in winners vs. losers; don't ask it to invent a strategy.
Skip Polymarket arbitrage prompts — Ron calls the fee math a non-starter and plans to rant about it in a separate video.

Video summary

Companion notes

Ron fixed OpenClaw's memory black box by piping daily conversations into Obsidian and backing them up to GitHub.

Why plain text beats vector memory

Ron's beef with OpenClaw's built-in "memory embeddings" and "memory indexing" is that they're a black box — "a bunch of numbers, a bunch of vector whatever" — and he can't audit what the agent actually retained. So at the end of each day he dumps the full chat into a markdown file, pushes it to an Obsidian vault, and mirrors it on GitHub. That gives him a readable journal and lets the agent learn over time, not just retrieve.

Two must-have Obsidian plugins

Smart Connections by Brian Petro — 836,000 downloads, runs a local embeddings model (no API key, private), does semantic + vector search over his vault.
QMD as MD — a dedicated plugin that forces the agent to write in a specific markdown format so context stays clean.

The Discord > WhatsApp argument

Ron's running his agent on Discord, not WhatsApp or Telegram, because Discord lets him run a topic per workflow. Two workflows live in there now:

1. Daily morning news briefing — cron job, fires 7 a.m. Hong Kong time, hits Brave Search API across X, crypto-specific sites, Finviz, TechCrunch, and the broader web, then writes a 2-minute structured summary to his Obsidian vault and appends any trading or video ideas to a projects ideas backlog. 2. Deep research — launches parallel sub-agents per topic. He's already noticed within two days that keeping contexts separated beats the "muddied" single-chat setup.

Where it's at, where it's going

He's two days in. Other users of the same Obsidian + GitHub combo reported "considerable improvement" and output that matches the author's voice by week 4–5. He wants the same — currently calling the news brief "pretty vanilla." Second goal: use the journal as a trading research assistant by feeding it his own trade log (entry, order flow, volume, delta) and letting the agent surface variables in winning vs. losing trades. He does not believe you can prompt your way to a profitable strategy, and he doesn't trust the Polymarket arbitrage hype on X because of the fee math.

Footgun to avoid

Set the cron timezone explicitly — "this was a hell of an annoying thing" when he didn't specify Hong Kong time and OpenClaw ran the job on UTC.

Watch on YouTube

Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=XfvzEqI4dkU