<aside> 🧠 Two prompts referenced in the video. Run Prompt 1 first to set up the memory system, then Prompt 2 to wire up the context infrastructure that feeds it. Works in Claude Code, Hermes, OpenClaw, Codex — any agent with file system access.
</aside>
Sets up the Obsidian-style brain with the raw → belief → memory pipeline from the video. Paste this into your agent first.
You are going to help me build a personal memory system for my AI agent,
modelled on an Obsidian-style linked knowledge graph with a raw → belief →
memory pipeline.
Set up the following file structure inside ~/Documents/agent-brain:
/raw → unprocessed dumps from internal + external sources
/wiki → synthesised, tagged memories (one .md per topic)
/output → finished deliverables the agent produces
log.md → chronological log of what happened across sessions
_index.md → master index of every memory in /wiki
Memory pipeline rules:
1. RAW LAYER
Anything new (call transcripts, social signals, articles, CRM updates,
business metrics) lands in /raw/YYYY-MM-DD-<source>-<slug>.md with full
source metadata: where it came from, when, why it was captured.
2. BELIEF CHECK
Before promoting raw data into a memory, check if the same signal has
appeared at least 3 times across /raw, OR if it's a black-and-white
factual statement I've explicitly told you (e.g. "we just hit $1M ARR").
One-off opinions do NOT become memories.
3. MEMORY LAYER
Promoted memories live in /wiki/<topic>.md with YAML frontmatter:
- tags: [list of semantic tags]
- sources: [list of /raw filenames that contributed]
- last_updated: ISO date
- confidence: low | medium | high
Body contains the synthesised insight in plain English, with backlinks
to related wiki entries using [[double-bracket]] syntax.
4. CRON SYNTHESIS
Once per day (or twice if I tell you), scan /raw for entries that have
not yet been processed, run the belief check, and either promote them
to /wiki or leave them in /raw with a "pending" note.
5. AGENT BEHAVIOUR
Before answering ANY question about my business, projects, preferences,
or strategy, read /wiki/_index.md and the last 20 entries of log.md.
The brain is canonical. Don't guess from your own context if the brain
has the answer.
Confirm the structure is set up, then show me the first /wiki/_index.md
and an empty log.md. Wait for me to start feeding you raw data before
populating anything else.
Wires up the internal + external context layers that feed the memory system from Prompt 1. Run this second.
You are going to help me wire up the full context infrastructure that
sits underneath my agent, mirroring the architecture from the video.
There are two context layers: INTERNAL and EXTERNAL. Both feed into the
memory system you've already set up (~/Documents/agent-brain).
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
INTERNAL CONTEXT (data I own about my business + life)
Connect or read from the following sources and dump raw exports into
~/Documents/agent-brain/raw/internal/:
• Notion → all docs, project pages, internal wikis
• HubSpot → contacts, deals, conversion ratios, pipeline value
• Slack → key channels I tell you to monitor
• Gmail → only threads I explicitly forward or label
• Granola → meeting notes + summaries
• Fireflies → call transcripts
• Featurebase → product feedback, changelogs, user requests
For each source, create a small ingestion script or MCP/connector config
that pulls new entries into /raw/internal/<source>/YYYY-MM-DD-<slug>.md
with full metadata at the top.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
EXTERNAL CONTEXT (signal from the outside world)
Use Trigify (https://trigify.io) as the unified API to pull social and
news signal into ~/Documents/agent-brain/raw/external/.
Configure two tiers:
Tier 1 — SIGNAL (high velocity, what's trending right now)
• X / Twitter
• LinkedIn
• TikTok
• Threads
• Instagram
Tier 2 — LEARNING (long-form, depth on a thesis)
• Reddit
• Substack
• YouTube
• Podcasts
Workflow:
1. Tier 1 surfaces trending topics relevant to my domain.
2. When a topic resonates (appears 3+ times or I flag it), automatically
trigger a Tier 2 deep dive on the same theme.
3. All raw pulls land in /raw/external/<platform>/YYYY-MM-DD-<slug>.md
with source URL, author, engagement metrics, and capture reason.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
DAILY LOOP
Set up a cron (or scheduled job in the agent) that:
1. Pulls latest from all internal + external sources into /raw.
2. Runs the belief check from the memory system prompt.
3. Promotes qualifying entries into /wiki.
4. Appends a one-line summary to log.md.
5. Reports back: what was ingested, what got promoted, what's still pending.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
SKILLS LAYER
Once the context + memory layer is alive, I'll attach skills on top:
AISDR, content writer, internal automations, etc. Every skill MUST read
from /wiki before producing output. No skill is allowed to generate
content from raw model knowledge alone.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Start by listing exactly which of these tools I have credentials for,
which connectors / MCPs you need me to install, and what's blocking the
first ingestion run. Don't build anything yet. Give me the gap analysis
first, then we work through it source by source.