-
-
Notifications
You must be signed in to change notification settings - Fork 39
Frequently Asked Questions
Common questions and honest answers about Second Brain.
Local files work well when your AI tools all share the same machine and you have a small number of notes. The tradeoff is that as your memory grows, you end up stuffing the entire context window with files hoping the relevant one is in there.
Second Brain retrieves only what’s relevant to the current conversation. Instead of loading everything, it queries by meaning and returns the most relevant memories. That keeps context usage low and scales much better as your knowledge base grows.
The other difference is cross-device. If you use Claude on your phone, in the browser, or on a different machine, local files don’t travel with you. Second Brain lives on Cloudflare so it’s available from anywhere.
If your tools all share the same local vault and you work from one machine, that setup is probably already working for you.
A virtual filesystem approach stores memories as files the AI reads directly. That works at small scale but has limits:
- As memory grows, you end up loading everything into the context window hoping the right file is in there
- There’s no way to retrieve by meaning — search is keyword or file-path based
- Context costs go up with every new memory you add
Embeddings let you retrieve only what’s relevant. Every note gets converted to a vector representing its meaning. When you query, the system finds the closest vectors by meaning, not by exact words. “Users dropping off at checkout” surfaces from a search for “onboarding problems” with no keyword overlap.
The tradeoff is that embeddings are more complex to set up and require an embedding model. Second Brain handles all of this automatically — you don’t need to think about it.
Claude’s built-in memory is a black box. You can’t search it, audit it, or control what gets stored. It occasionally surfaces something useful but you have no visibility into what it holds or why.
Second Brain gives you full control. You can see every memory in the dashboard, search across all of them, append updates, and delete anything you didn’t intend to keep. Nothing gets stored without going through the same endpoints you control.
It also works across AI tools. Claude’s built-in memory only works in Claude. Second Brain works with Claude, ChatGPT, Cursor, and any MCP-compatible client.
No. The recall call at conversation start adds maybe 500-2000 tokens depending on how many entries come back — negligible against a 200k context window. You can also tune topK down if you want to minimize it.
Store calls are tiny, effectively free.
The only real cost is if Claude is storing its own responses after every turn. The fix is to only store responses that contain something genuinely reusable — a decision, a recommendation, a solution — not every reply.
Net effect for most conversations: neutral to slightly positive on context efficiency. Instead of re-pasting context manually at the start of every session, recall does it in one targeted call.
Real problem. If you store a decision and later change course without updating the memory, Claude might confidently surface the outdated version.
The fix: use append when something changes rather than storing a new conflicting entry — it adds a timestamped update to the existing entry, so the history is preserved and the most recent state is clear. Use update instead if the old content should be replaced entirely rather than appended to.
For anything that’s completely outdated, use forget to remove it entirely.
You also don't need to manage this by hand in most cases — recall already applies time-decay reranking, which automatically deprioritizes older memories so the freshest relevant version surfaces first (see How It Works for the details). remember also runs contradiction detection and smart merging, so storing an updated version of something often resolves the conflict for you automatically.
Every note gets embedded as a 384-dimensional vector using bge-small-en-v1.5 on Workers AI. When you call recall, your query gets embedded the same way and Cloudflare Vectorize finds the closest stored vectors by cosine similarity.
This means recall works by meaning, not keywords. You don’t need to remember exactly how you phrased something when you saved it.
Yes. Community members have already swapped in Qwen with 1024 dimensions for better multilingual performance. To change the model, update the embedding model name in the Worker and update the Vectorize index dimensions to match.
Everything lives in your own Cloudflare account — your D1 database and your Vectorize index. No third party ever touches your memories after deployment. Cloudflare can access your infrastructure as your cloud provider, but Second Brain as a project has zero visibility into what you store.
Yes. Second Brain is an MCP server. Any MCP-compatible client can connect to it — Claude Desktop, Claude Code, claude.ai, Cursor, ChatGPT (in beta via Developer Mode for Plus/Pro users), and others.
The Obsidian plugin syncs your vault notes directly to Second Brain, making them available to any of these clients.
It’s stable for personal use. The one-click Deploy to Cloudflare button provisions everything — D1, Vectorize, KV, and the schema — and prompts you to set your AUTH_TOKEN during deploy, so setup is genuinely one-click for most people. You'll still want basic familiarity with Cloudflare Workers if something needs troubleshooting.
For production workloads with multiple users, you’d want to think carefully about the write policy and access controls. The current design is built for personal, single-user use.
Open a Discussion — not an Issue — for feature requests and ideas. Issues are for bugs with steps to reproduce. Discussions is also where releases get announced and the roadmap gets shaped.