Hearing terms like "agent inbox" and "memory layer" and wondering what they mean in practice? Here's the working vocabulary of AI-native productivity, defined plainly, with an honest note on where a hub like Notestream fits.
New categories breed new jargon, and the personal AI space has bred plenty of it. The good news is that most of these terms describe real layers of a working setup, not marketing inventions. Learn eight of them and every product page in the category becomes easier to read. Definitions first, architecture second.
The glossary
Capture surface
Any entry point where a thought can leave your head and enter your system: a text field, a forwarded email, an API call. The quality bar is simple. Capture should take seconds, and it should be available from wherever you are already working. In Notestream, the capture surfaces are the assistant field in the web app, email forwarding, and the MCP API, all feeding one pipeline.
One detail matters more than the count of surfaces: what travels through them. A capture is a focused thought (an idea, a commitment, a deadline), never a meeting-sized document. You type or forward the thought as it occurs, and the system takes it from there.
Classification on arrival
The practice of filing each item the moment it is captured, rather than in a later sorting session. The AI reads your capture as you submit it and decides two things: what it is (a task, a sub-task, a reminder, a reference note) and where it belongs (which project). There is no pile to sort on Friday, because nothing waits in a pile.
Contrast this with batch sorting, where captures accumulate in an unsorted inbox until you spend an evening triaging. Batch sorting is where most personal systems quietly stall. Classification on arrival removes the triage session by doing the filing at submit time, one capture at a time.
Memory layer
A durable record of your decisions, commitments, and knowledge that lives outside any single AI model. Because it sits outside the models, it survives tool switches and serves every model you use. This is distinct from the built-in memory that ChatGPT, Claude, and Gemini each offer, which is useful for conversational continuity but serves only that one model and stays behind when you move.
A good memory layer can be questioned in plain language. Ask it what you decided about the vendor, and you get a sourced answer, not a search results page. If you want the deeper version of this idea, we wrote about building a searchable memory for your AI work.
Agent inbox
The single place where results from your AI tools land for review. Think of the inbox pattern applied to agent output: everything arrives in one stream, each item gets a decision (approve, revise, or discard), and nothing silently accumulates. Without an agent inbox, output scatters across each tool's own history, and "what did my AI finish this week?" becomes a twenty-minute archaeology project.
Dispatch hub
The layer from which work is routed out to AI tools: this bug to a coding agent, that draft to a writing model, this design idea to a prototyping tool. Dispatching from a hub, rather than opening ad-hoc sessions, is what makes AI work trackable. The assignment, the status, and the result all live in the record instead of in your head.
Intent routing
Inside an assistant, the step that decides what kind of request you made and which machinery should answer it. "How many open tasks do I have?" needs a database query that counts everything. "What did I decide about the vendor?" needs meaning-based search over your notes, with sources cited. "Remind me Thursday" needs a proposed change that you confirm before it lands. One question box, three different engines behind it.
Thread-scoped retrieval
Search that respects project boundaries. A question asked inside a project searches that project's record by default and goes account-wide only when you ask. It sounds minor until you have six projects with overlapping vocabulary. Then it is the difference between an answer and a noise pile.
Human-in-the-loop review
The rule that AI-proposed changes and AI-produced results get explicit approval before they merge into your record. This is the least glamorous term in the glossary and the one that makes the rest trustworthy. A system that writes to your record without review is a system you will eventually stop checking.
How the layers stack
Read the glossary bottom to top and you get the architecture of an AI-native day. Capture surfaces feed classification. Classification builds the memory layer. The memory layer feeds intent routing, so you can question everything you've recorded. The dispatch hub sends work out to your models, and the agent inbox catches what comes back, gated by review.
One property is worth noticing: every layer is model-agnostic. Whichever of Claude, ChatGPT, or Gemini is ahead this quarter, the layers stay the same. Models churn. The layers persist. That is why the vocabulary is worth learning even if you never adopt a single new tool.
Where the terms come from
None of these ideas appeared from nowhere, and the tools you already know each solved part of the puzzle. The agent inbox borrows decades of email thinking. The memory layer echoes what Obsidian and Notion have long done well: a durable personal record you own and shape. NotebookLM is excellent when the job is questioning a fixed set of documents. Model-side memory in ChatGPT, Claude, and Gemini keeps individual conversations coherent across sessions.
Each of those tools earns its place. What the newer vocabulary names is the connective tissue among them: one record that feeds capture, questioning, and dispatch at the same time, for the person doing the work rather than the software vendor describing it.
Where Notestream sits: a central hub for your AI workflow
Notestream is these layers assembled as one product for an individual operator. Three capture surfaces (the assistant field, forwarded email, the MCP API). Classification on arrival, handled automatically by Notestream's AI. A memory you can question inside each project.
On the outbound side: intent routing behind one assistant box, dispatch to Claude, Claude Code, Lovable, or Claude Cowork, plus in-app chat with Claude, ChatGPT, or Gemini. And a review queue where every result lands for your approval.
Could you assemble the same stack from parts? You could, and for some setups that is the right call. A dedicated outliner will beat a hub on outlining depth, and a dedicated research tool will beat it on document analysis. The point of a hub is that the layers work because they share one record. Nothing needs syncing, and nothing falls between tools.
The vocabulary in action
Here is one afternoon, translated. At 2:15 you notice the onboarding email still links to the old pricing page, so you type one line into the assistant field: "Onboarding email links to the old pricing page, fix the template." Classification on arrival files it as a task in your marketing project. At 2:40 a user email sparks a settings-screen idea, so you forward it to your capture address, and it lands as a task in the product project.
Then you dispatch. The template fix goes to Claude Code with a one-line note about where the file lives. The settings-screen idea goes to Lovable for a first visual pass. While those run, you ask the in-app assistant, "what's still open in marketing this week?" and intent routing answers with a complete count, not a guess.
An hour later both results are sitting in your agent inbox. You approve the template fix, send the settings screen back with one revision note, and the record updates itself. That was a capture surface, classification on arrival, a dispatch hub, and human-in-the-loop review doing their jobs in a single afternoon, with two different AI tools working from one hub.
See the vocabulary in action on your own stack. Try it free at notestream.ai.