Second Brain for AI
Persistent memory for Claude, ChatGPT & Cursor. Free.
Start new thread

Second Brain for Mac and Windows - Your AI memory, everywhere you work

by
•
The Second Brain desktop app gets you running in five minutes. No terminal. No git. Just download, set a password, and sign into Cloudflare. Then the real magic starts. Persistent memory across every AI tool you use. Recall that just got a lot smarter. Surfaces the right memory, not just the newest one. A knowledge graph that builds itself. Your data in your own Cloudflare account, not ours. Mac and Windows. You asked for the app. You asked for better recall. We shipped both.

Add a comment

Replies

Best
🧐 Good find

The override versus accumulate question above is the real one. Semantic search alone can't answer it either. An old architecture decision and its replacement will usually score just as similar to each other as either does to the query looking for it. That has to get resolved somewhere: at write time, checking new entries against existing ones and versioning what's superseded, or at read time, dumping both into context and trusting the model to notice the contradiction. Which is it here?

Maker
Write time. Each new memory is checked against its nearest existing ones as it lands, and a model call, not the similarity score, decides whether it’s a real contradiction, an update, or just adjacent. Semantic search only shortlists the candidates. The outcome then gets written down. Conflict with a canonical entry: the incumbent stays authoritative and the newcomer is kept as a draft. Conflict with a non-canonical one: the new entry wins, the old is deprecated (kept for audit, out of recall), and a “supersedes” edge records the replacement. So it’s resolved and versioned before anything is read back. Read time isn’t asked to notice the contradiction.

 That's an actual mechanism, not a hand wave. The canonical versus non-canonical split is what I'd push on next. Something has to decide an entry is canonical in the first place. Is that promotion automatic after enough confirmations, or does a human have to mark it. And once something is canonical, can it ever get knocked back down, or does canonical status stick even after the model call finds a real contradiction against it.

Maker

 Good question, and it’s not a confirmation counter either way.

Two things can make an entry canonical. When a memory gets written, a model call marks it canonical only if it’s a confirmed decision, a durable fact, or a stated permanent preference, and it’s tuned to be strict. We judge the content instead of counting confirmations on purpose: repetition isn’t truth. Something said five times isn’t more authoritative than something said once, so a counter would just promote whatever gets repeated most. You can also mark it yourself with set_status, and the auto step only runs when an entry has no status yet, so it never overrides you. Your call always beats the model’s guess.

Contradictions don’t knock it down. When a new write clashes with a canonical entry, the canonical one wins and the newcomer lands as a draft. That’s the point. The failure mode we were fixing is recency winning by default, where the newest write silently overwrites a settled decision and you get a confidently wrong answer later. Most writes that contradict a settled fact are noise or missing context, not a real reversal, so a contradiction surfaces both sides and flags it for you instead of flipping authority on its own. The one thing that demotes canonical is you calling set_status. Protected by default, reversible when you mean it. The model can nominate, but reversing settled truth stays your call, because that’s the one decision you don’t want automated

First
Previous
123