Sinapsis

A personal knowledge base powered by LLMs and Obsidian. Instead of searching raw documents every time (RAG), an LLM incrementally builds and maintains a structured, interlinked wiki that compounds over time.

Based on Andrej Karpathy's LLM Wiki pattern.

How it works

You capture raw sources throughout your day — meeting notes, articles, Slack threads, documents. When you're ready, you tell the LLM to ingest them. It reads each source, extracts key information, and integrates it into the wiki: creating entity pages, updating timelines, noting decisions, cross-referencing everything. The knowledge is compiled once and kept current.

Works with Claude Code and Gemini CLI. The schema file (CLAUDE.md / GEMINI.md) is symlinked so both stay in sync.

You (capture) → raw/ → LLM (ingest) → wiki/ → You (query, browse, generate)

Three layers:

Layer	What	Who owns it
raw/	Source documents (meeting notes, articles, docs, Slack)	You — immutable
wiki/	Structured, interlinked markdown pages	LLM — auto-maintained
CLAUDE.md / GEMINI.md	Schema and conventions governing LLM behavior	Co-evolved

Getting started

Prerequisites

• Obsidian (free)
• An LLM CLI: Claude Code or Gemini CLI
• Git

Setup

1. Clone the repo

   git clone https://github.com/antonioaragues/sinapsis.git

2. Open as Obsidian vault

   Open Obsidian → "Open folder as vault" → select sinapsis/.

3. Install community plugins

   Go to Settings → Community plugins → Enable community plugins, then install:

   • Templater — for source capture templates
   • Dataview — for dynamic queries over wiki pages
   • Marp Slides — for generating presentations (optional)

4. Configure Templater

   Settings → Templater → Template folder location: templates/

5. Configure Daily Notes (core plugin)

   Settings → Daily notes:

   • New file location: raw/daily
   • Date format: YYYY-MM-DD
   • Template file location: templates/daily.md

6. Install Tasks plugin (recommended for action items)

   Settings → Community plugins → install Tasks. The wiki extracts action items from your sources into [[wiki/action-items.md]] using Tasks plugin syntax.

7. Done. Open your LLM CLI (claude or gemini) in the sinapsis/ directory and start ingesting.

Daily workflow

1. Capture (throughout the day)

Use the daily note as your scratchpad. Open it with the daily note button or hotkey.

# 2026-04-14, Monday

## Meetings
### Standup
- Feature X delayed, dependency on Platform team
- @maria raised concern about API rate limits

### 1:1 with James (VP Eng)
- Migration timeline: wants Q3, not Q4
- Decision: prioritize auth refactor over new dashboard

## Slack
- #product-strategy: competitor Y launching process mining for SAP

## Thoughts
- Revisit scoring model after Q1 data

For other sources:

• Web articles: use Obsidian Web Clipper → saves to raw/articles/
• Meeting transcripts: paste into raw/meetings/ using the meeting template (Templater)
• Slack threads: copy-paste into raw/slack/ using the Slack template
• Google Docs: ingested directly from Drive — no export needed (see Google Docs integration)

All templates are available via Templater (Ctrl/Cmd + T or the command palette).

2. Update (end of day or whenever)

Open your LLM CLI in the sinapsis/ directory:

# The all-in-one command — processes everything new in a single pass
> update everything

This does two things:

1. Finds new local files in raw/ that haven't been ingested yet
2. Re-syncs all tracked Google Docs for changes

You can also be more specific:

# Only new local files
> ingest all new

# Only today's captures
> ingest today

# Only a specific file
> ingest raw/articles/some-article.md

# Only a specific folder
> ingest all new meetings

# Only Google Docs sync
> sync google docs

The LLM will:

• Read each source
• Create/update entity, concept, decision, and comparison pages
• Maintain cross-references and timelines
• Update the index and log
• Give you a summary of everything that changed

3. Query (anytime)

Ask questions against the wiki:

> What decisions have been made about the auth refactor?
> How does our process mining approach compare to competitor Y?
> What AI trends are relevant to our roadmap?
> Give me a briefing on everything discussed about Feature X in the last 2 weeks

The LLM searches the wiki index, reads relevant pages, and synthesizes an answer with citations. Good answers can be filed back into the wiki as new pages.

4. Generate (on demand)

Create artifacts from wiki content:

> Create a briefing on the Q3 migration plan
> Generate slides summarizing our competitive position

Outputs go to outputs/briefings/ or outputs/slides/ (Marp format).

5. Lint (periodically)

Ask the LLM to health-check the wiki:

> lint the wiki

Finds contradictions, orphan pages, stale claims, missing cross-references, and gaps.

Slash commands

Common operations are available as Claude Code slash commands (in .claude/commands/). Type / in Claude Code to see the list:

Command	What it does
/update	Process all new local files + sync Google Docs in one pass
/ingest-today	Ingest today's daily note and any other files modified today
/ingest-new	Ingest all un-ingested local files in raw/
/sync-docs	Sync all tracked Google Docs
/lint	Health-check the wiki for contradictions, orphans, gaps
/briefing <topic>	Generate a briefing on a topic
/pre-meeting <name>	Pre-meeting briefing with context, recent activity, talking points
/weekly-review	Weekly summary of activity, decisions, open threads
/action-items	Show all open action items grouped by due date

Action items

Action items are extracted from sources during ingest and written in Obsidian Tasks plugin format. They aggregate automatically in [[wiki/action-items.md]].

Format example:

- [ ] Prepare Q3 roadmap deck for board 📅 2026-04-25 👤 @antonio 🔼
- [x] Review auth refactor RFC ✅ 2026-04-12 👤 @james

The Tasks plugin reads these from anywhere in the vault — source summaries, entity pages, daily notes — and the queries in wiki/action-items.md show overdue, this-week, later, and recently-completed views.

Use /action-items for a summary across the wiki, or open wiki/action-items.md in Obsidian to interact with tasks (check them off, edit due dates, etc.).

Scheduled sync

Run update everything automatically every weekday at 22:00 (configurable).

Install (macOS, launchd):

./scripts/install-scheduled-sync.sh

This installs a launchd agent that runs scripts/scheduled-sync.sh Mon-Fri at 22:00. The script:

1. Runs /update non-interactively (Claude Code with --permission-mode acceptEdits)
2. Auto-commits any wiki changes to git
3. Pushes to remote
4. Logs everything to .sync-logs/

To run manually:

./scripts/scheduled-sync.sh           # run now
launchctl start com.sinapsis.scheduled-sync   # trigger via launchd

To uninstall:

launchctl unload ~/Library/LaunchAgents/com.sinapsis.scheduled-sync.plist
rm ~/Library/LaunchAgents/com.sinapsis.scheduled-sync.plist

To change schedule — edit scripts/com.sinapsis.scheduled-sync.plist and reinstall.

Google Docs integration

Google Docs are "living" documents that change over time. Instead of manually exporting them, Sinapsis reads them directly from Google Drive and tracks changes automatically.

Setup

1. Connect Google Drive to your LLM CLI:

   Claude Code — add the Google Drive MCP server to your settings (~/.claude/settings.json):

   {
     "mcpServers": {
       "google-drive": {
         "type": "url",
         "url": "https://mcp.google.com/gdrive"
       }
     }
   }

   Gemini CLI — Google Drive access is built-in. Enable it with the @google-drive extension in your Gemini settings.

2. Authenticate — the first time the LLM accesses Drive, it will prompt you to authorize via OAuth in the browser.

3. Done. You can now ingest Google Docs directly by URL or search for them from Drive.

Usage

Ingest a Google Doc by URL:

> ingest https://docs.google.com/document/d/1abc123.../edit

The LLM reads the doc directly from Drive, creates a local snapshot in raw/docs/, processes it through the standard ingest workflow, and registers it for future syncing.

Search your Drive:

> search Drive for Q3 roadmap
> find docs modified this week

The LLM searches your Google Drive and presents results. You choose which ones to ingest.

Sync tracked docs:

Once a Google Doc is ingested, it's registered in wiki/gdoc-registry.md. Sync pulls the latest version and processes any changes:

# Sync all tracked Google Docs
> sync google docs

# Sync a specific doc
> sync Q3 Roadmap PRD

The LLM will:

• Re-fetch each registered doc from Drive
• Compare with the previous snapshot
• Update source summaries and wiki pages with what changed
• Add timeline entries noting the evolution
• Skip docs that haven't changed

This is ideal for living documents like PRDs, design docs, and roadmaps that evolve over weeks or months. The wiki tracks the full history of changes.

Directory structure

sinapsis/
├── CLAUDE.md              # Schema — LLM behavior rules and conventions (GEMINI.md is a symlink)
├── .claude/commands/      # Custom slash commands for Claude Code
├── scripts/               # Scheduled sync installer + script
├── raw/                   # Your source documents (immutable)
│   ├── daily/             # Daily notes (scratchpad)
│   ├── articles/          # Web clips
│   ├── meetings/          # Transcriptions
│   ├── docs/              # PRDs, design docs, Google Docs
│   ├── slack/             # Slack exports
│   ├── feedback/          # Customer feedback
│   └── assets/            # Images, attachments
├── wiki/                  # LLM-maintained knowledge base
│   ├── index.md           # Master page index
│   ├── log.md             # Activity log (chronological)
│   ├── overview.md        # High-level synthesis
│   ├── action-items.md    # Aggregated action items (Tasks plugin)
│   ├── gdoc-registry.md   # Tracked Google Docs for sync
│   ├── entities/          # Products, features, people, teams, companies
│   ├── concepts/          # Technical concepts, methodologies
│   ├── decisions/         # Decision records with timelines
│   ├── comparisons/       # Competitive analysis, trade-offs
│   └── sources/           # One summary per ingested source
├── outputs/               # Generated artifacts
│   ├── briefings/
│   └── slides/
└── templates/             # Templater templates for quick capture
    ├── daily.md
    ├── meeting.md
    ├── article.md
    ├── doc.md
    ├── slack-thread.md
    └── feedback.md

Wiki page types

Type	Location	Purpose
Entity	wiki/entities/	Products, features, people, teams, companies. Includes timeline.
Concept	wiki/concepts/	Technical concepts, methodologies, patterns. Includes relevance analysis.
Decision	wiki/decisions/	Decision records with context, options, outcome, and full timeline.
Comparison	wiki/comparisons/	Side-by-side analysis with comparison tables.
Source	wiki/sources/	Summary of each ingested raw source with key takeaways.

All pages use YAML frontmatter (queryable with Dataview) and [[wikilinks]] for cross-referencing (visible in Obsidian's graph view).

Customization

The system is designed to be adapted. Key configuration:

• CLAUDE.md (symlinked as GEMINI.md) — the main schema. Modify domains, page templates, workflows, writing style, naming conventions. This is the most important file. Edit CLAUDE.md and the symlink keeps GEMINI.md in sync automatically.
• templates/ — Templater templates for raw source capture. Add new ones for source types specific to your workflow.
• .obsidian/ — Obsidian settings. Adjust to your preferences.

Inspired by

• LLM Wiki by Andrej Karpathy
• Memex by Vannevar Bush (1945)