Vault

Save conversation context to an Obsidian vault. The model infers what to save, how to categorize it, and where it goes. The user never needs to remember commands.

How It Works

1. User signals intent to save (or agent suggests it after a good session)
2. Agent extracts the relevant content from conversation
3. Agent categorizes it and picks the right folder
4. Agent composes markdown with frontmatter
5. vault.sh writes the file and auto git-commits

Categories

Infer the category from context. When ambiguous, ask.

Category	Folder	When to use
Session dump	daily/	End of a work session — capture decisions, takeaways, open threads
Decision	decisions/	A choice was made with reasoning — "we're going with X because Y"
Lesson	learning/	An insight, pattern, or thing learned — "TIL," "turns out," aha moments
Idea	ideas/	Something to explore later — side projects, features, experiments
Reference	learning/	A link, article, tweet, or tool worth retrieving later

Using vault.sh

The script handles file creation, frontmatter, and git-commit. Pipe markdown content to it:

echo '<markdown content>' | bash scripts/vault.sh <command> "title"

Commands: done, decide, learn, idea, ref, search, sync

• done [topic] — Appends to today's daily/YYYY-MM-DD.md
• decide "title" — Creates decisions/<slug>.md
• learn "title" — Creates learning/<slug>.md
• idea "title" — Creates ideas/<slug>.md
• ref "title" — Creates learning/<slug>.md (tagged reference)
• search "query" — Grep vault content
• sync — Commit + push to remote

Content Structure

Adapt the structure to the category:

Session dump:

## Topic Name

### Decisions
- What was decided and why

### Key Context
- Important things discussed

### Open Threads
- [ ] What's unfinished

### References
- Links, tickets, people involved

Decision:

**Decision:** What was decided
**Context:** Why this came up
**Reasoning:** Why this over alternatives
**Related:** Tickets, links, people

Lesson:

**Insight:** The core takeaway
**Context:** How we learned this
**Source:** Where it came from

Idea: Keep lightweight — one-liner + why it's interesting + next step.

Reference: URL + summary + key takeaways.

Frontmatter

Every entry gets:

---
title: "Title"
created: YYYY-MM-DD
tags: [type, topic1, topic2]
related: [DES-123, link, etc]  # when applicable
---

Use consistent tags:

• Type: daily, decision, learning, idea, reference
• Source: telegram, claude-code, call, article
• Topic: whatever fits — project-x, infra, design, etc.

Telegram Behavior

Telegram is one long conversation. When inferring what to save:

1. If the user says "save this" about a specific thing — capture that thing
2. If the user says "done" or "wrap up" — capture the full session since the last natural break
3. If multiple topics were discussed and it's ambiguous — ask: "Which part? The X discussion, the Y brainstorm, or everything?"
4. Skip heartbeat messages — they're noise

Retrieval

When the user asks "what did we decide about X" or "find that thing about Y":

bash scripts/vault.sh search "query"

Then read the matching files and summarize.

Proactive Suggestions

After a meaningful session, suggest saving. Keep it casual:

• "Want me to save this session to the vault?"
• "Good stuff — should I vault this?"

Don't suggest after casual chat, quick questions, or quiet periods.