Memory Mason
Capture AI conversations. Build an Obsidian knowledge base. One command.
What It Does
Memory Mason hooks into your AI coding agent and silently captures every conversation into daily Obsidian logs. When you're ready, run /mmc to compile raw logs into structured knowledge articles — concepts, synthesis pages, and MOCs — all interlinked with [[wikilinks]]. Then use /mmq to retrieve answers from your compiled knowledge base without leaving the conversation.
No API key needed. No cloud sync. Everything stays local in your Obsidian vault.
How data flows in
[AI Conversation] ──> [Hook Runtime] ──> [Smart Filter] ──> [Obsidian Vault]
(any agent) (automatic) (6-stage pipe) _raw/YYYY-MM-DD/
Hooks capture prompts, tool results, and session transcripts into daily log files. Before anything reaches your vault, a six-stage filtering pipeline strips noise and protects sensitive data:
- Tag stripping — Removes system-reminder, system-instruction, and other injected tags
- ANSI removal — Strips terminal control characters from tool output
- Event classification — Categorizes each event as error, test result, discovery, exploration, meta, or noise. Exploration and meta events are discarded; errors and test results are always kept.
- Prose compression — Removes filler words and hedging phrases while preserving code blocks, URLs, inline code, and quoted strings
- Sensitive content blocking — Skips capture when input contains credentials, private keys,
.env contents, or paths like .ssh/ and .aws/
- Deduplication — Content-hash check prevents the same data from being written twice within a session
Memory Mason's own commands (/mmc, /mmq, /mml, /mms, /mma, /mmsetup) and namespaced /memory-mason:* forms are automatically excluded from capture.
Daily logs are stored in per-day folders and auto-split into files of up to 500KB each. This keeps Obsidian responsive and ensures each file stays within LLM processing limits. No data is lost — every conversation turn is preserved, and Obsidian indexes all chunks for full-text search. Token economics (raw vs. stored token counts and savings percentage) are tracked automatically and reported by /mms. See docs/README.md for technical details on chunked storage.
How knowledge is built
_raw/YYYY-MM-DD/ ──> /mmc compile ──> concepts/
atlas/
synthesis/
index.md
Run /mmc to compile daily logs into structured articles. The host LLM reads your raw logs and produces atomic concept pages, MOC navigation pages in atlas/, and cross-session synthesis pages — all linked with [[wikilinks]] for Obsidian graph navigation.
How knowledge is retrieved
/mmq "How does auth work?" ──> _meta/context.md ──> concepts/ + atlas/ + synthesis/ ──> answer with [[citations]]
Run /mmq with a question. Memory Mason checks session context for recent focus first, then reads compiled articles, synthesizes an answer, and cites sources with [[wikilinks]] back to the original concepts. Your knowledge base grows with every session and becomes more useful over time.
Commands
| Command | What it does |
|---|
/mmc | Compile raw captures into concepts, MOCs, and synthesis pages; update session context and source manifest |
/mmq | Answer questions from your knowledge base with source citations |
/mml | Run knowledge base health checks (broken links, stale content, manifest integrity, and more) |
/mms | Show knowledge base status, token economics, health summary, and compilation coverage |
/mma | Archive old build log entries to keep the knowledge base log compact |
/mmsetup | First-time vault configuration (or uninstall) |
Install
Pick your agent. Restart the host after install.
rohitg00
thedotmack