local-first · open source · your disk
second-brain captures every AI session into Markdown on your machine. Switch models, clear the window, close the laptop. Nothing walks away.
Memory is Markdown under your home folder. Export is a folder copy.
Claude, GPT, Gemini, Grok. The brain does not move when the chat does.
Recall only what matters. Skip the slow compaction tax on every long thread.
works with
The vault is plain files, so any model can read it now. Automatic capture and recall land in the other apps through a small MCP server we're building next.
The context tax
Long chats re-send their history on every turn. Past half full, you pay for bloat you already know. second-brain puts the hard parts in Markdown so the vault grows and the window stays light.
scroll to play
Without second-brain
nothing on disk — all in the window
extra tokens re-carried this session
0
History stays in the chat. Nothing hits disk.
With second-brain
waiting for capture…
extra tokens above 50% line
0
Work happens. Notes land in Obsidian. Session stays thin.
Scroll to run the loop. Left fills the window. Right writes the vault and stays under the tax line.
Past 50% on a 1M window
On a 1M-token window, everything above 50% is optional bloat you re-send each turn. Math: extra = 1M × (usage% − 50%). Prices use Claude Opus 4.8 input at $5/MTok. Output cost is separate.
at 90% context
400k
extra tokens re-sent every query
extra cost / query
$2
Opus 4.8 · $5/MTok input only
saved if you stay ≤50%
$80/day
40 queries/day · 90% → ≤50%
| Context used | Extra tokens / query | Extra $ / query | Extra $ / day40 queries | Kept with second-brainstay ≤50% · /mo (22d) |
|---|---|---|---|---|
| 60%(600k carried) | +100,000 | +$0.50 | +$20 | $440 /mo |
| 70%(700k carried) | +200,000 | +$1 | +$40 | $880 /mo |
| 80%(800k carried) | +300,000 | +$1.50 | +$60 | $1,320 /mo |
| 90%(900k carried) | +400,000 | +$2 | +$80 | $1,760 /mo |
| 100%(1000k carried) | +500,000 | +$2.50 | +$100 | $2,200 /mo |
How to read this. At 90% you re-send 900k tokens every query. Only 500k sits under the 50% line — the other 400k is pure tax ($2 / query). second-brain keeps durable memory on disk and recalls a small slice, so you can /clear and live under 50%. “Saved” is that input tax going to zero — not free tokens, just not re-paying Opus for history you already wrote down.
why it exists
Vendor memory is a moat. second-brain is a folder. That is the whole product thesis, stated plainly.
The retry params you set on Tuesday, the bug you fixed last session, the decision you already paid for. It lands in Obsidian and stays. Clear the chat; the vault still knows.
One linked vault, not a silo per repo. What you learned shipping A shows up when you start B. Skills stack across projects instead of dying in dead chats.
Plain files on your disk. Jump between ChatGPT, Claude, Gemini, and Grok. The memory follows you because it was never trapped in a vendor.
Wikilinked .md files. Offline, git-versioned, no server, no login. Uninstall the tool tomorrow and every note remains yours.
Memory lives outside the window. Recall only what matters for this prompt, and skip the slow compaction that burns tokens to remember badly.
Obsidian-compatible notes, captured as you work. The vault you meant to keep by hand, building on its own. Always remembering.
Local, file-based, cross-model, hands-free. Most tools manage one.
swipe the table →
| second-brain | ChatGPT memory | Claude memory | mem0 | Basic Memory | |
|---|---|---|---|---|---|
| Where it's stored | your disk | OpenAI cloud | Anthropic cloud | DB / cloud | your disk |
| You own & export it | ✓ | – | limited | self-host | ✓ |
| Works across models | ✓ | – | – | ✓ | ✓ |
| Auto-captures sessions | ✓ | auto | auto | – | – |
| No account, no server | ✓ | – | – | – | ✓ |
Where peers beat us: Basic Memory and Pieces add semantic search; mem0 adds knowledge-graph retrieval at agent scale. second-brain trades that for zero-infra, no-account ownership and native Claude Code fit: grep and plain Markdown over a database.