Ingest-Route-Lint AI Second Brain
Continuously turn scattered company activity into organized agent-ready memory
- Difficulty
- Advanced
- Time to result
- ~weeks to results
- Steps
- 7
- Confidence
- 99%
Create a local LLM-friendly wiki whose structure reflects what the organization actually manages. New documents, meeting transcripts, and communications first enter a raw staging folder. An ingest command checks the previous run time, processes only the newer material, extracts relevant intelligence, and routes it into primary categories such as strategic bets and people. Important concepts that do not fit those categories receive their own wiki pages, while repeatedly encountered people gain dedicated files. A separate lint command examines the knowledge layer for contradictions, duplication, and cleanup decisions. Finally, the vault syncs across authorized environments so different agent interfaces use the same memory. This pipeline turns scattered activity into durable context while preserving a review mechanism for conflicting knowledge.
Origin
Extracted from Marketing Against The Grain when the hosts compared their Obsidian-based company intelligence systems.
Core principles
- 01Raw information needs a staging area before synthesis
- 02New information should be processed incrementally
- 03Knowledge should map to the outcomes and people that matter
- 04Repeated uncategorized concepts deserve their own pages
- 05Contradictions require deliberate reconciliation
- 06The agent should know where to retrieve each class of intelligence
How to run it
- 1
Stage Raw Intelligence
Place incoming documents, transcripts, and communication exports into a designated raw folder before synthesis.
Pro tip Automate repeatable inputs such as converting shared video recordings into transcripts.
Watch out Do not mix unprocessed source material directly into curated knowledge pages.
- 2
Track the Ingestion Cursor
Record when the last successful ingest occurred so each run can focus on the new interval.
Pro tip Use a log file that the agent can inspect before collecting updates.
Watch out A missing cursor can cause repeated processing or gaps.
- 3
Extract Relevant Intelligence
Define the types of facts, decisions, blockers, and changes the system should look for across the new material.
Pro tip Tie extraction criteria to actual management questions and outcomes.
Watch out Collecting everything without priorities creates a noisy vault.
- 4
Route Into Core Domains
Assign extracted intelligence to stable domains such as strategic bets, projects, teams, or direct reports.
Pro tip Use categories that mirror how decisions are made, not merely the source application.
Watch out Source-based filing alone can fragment related intelligence.
- 5
Promote Emergent Topics
Create a wiki page when a meaningful concept or person appears repeatedly but does not fit the primary structure.
Pro tip Use recurrence thresholds, such as a person appearing two or three times, to avoid premature page creation.
Watch out Creating a page for every mention produces clutter.
- 6
Lint the Knowledge Layer
Search for contradictions, stale claims, duplication, and organizational problems, then request a decision where reconciliation is ambiguous.
Pro tip Preserve links to source material when resolving disputed information.
Watch out Do not let the agent silently choose between consequential conflicting claims.
- 7
Sync the Vault
Synchronize the curated vault across the approved computers and agent interfaces that need shared context.
Pro tip Use version control or managed Obsidian sync depending on team requirements.
Watch out Protect sensitive company knowledge and avoid syncing it to unauthorized environments.
In the wild
A meeting recorder captures sessions the leader misses and places transcripts into the raw staging area. On the next ingest, the system extracts team blockers, updates people and strategic-bet files, and creates pages for recurring concepts. When the leader asks for a status update, the agent combines Slack, meetings, goals, and project context into a focused answer.
→ The leader receives current, decision-relevant intelligence without manually reviewing every source.
Common mistakes
Ingesting Without a Cursor
Reprocessing the entire history wastes time and increases the chance of duplicated or conflicting notes.
Organizing Only by Source
Separating Slack, meetings, and documents prevents the agent from seeing one outcome or person across channels.
Skipping Contradiction Checks
An accumulating vault can retain incompatible claims unless a lint process identifies and reconciles them.
Is it for you?
Best for
Leaders and teams whose decisions depend on information scattered across meetings, documents, communications, projects, and people.
Not ideal for
Small or static information sets that can be maintained accurately in a single concise context file.
From the transcript
“I have a staging folder called raw, and so I download all of my docs that I get. I have a script that will basically…”
“And then I have logic that will say these are the types of intelligence I'm looking for, and it will route it to different places.”
“And then I have a command called lint that will look across and see and try to clean it up. So we'll say, hey, this…”
From the episode
Build An Army Of AI Employees To Run Your Business (With Hermes AI)