Building tools for People who build things

Right now that’s Droplet, a desktop AI assistant that reads real papers and cites its sources.

Install Droplet Get in touch

macOS and Windows · v1.0.0

Newsletter

What’s next

We’re starting with Droplet as a precision desktop coding tool.

Our longer-term work focuses on spatial software models and autonomous environment reasoning engines.

semantic documentation research

YOU

Research how long-term memory is being implemented in modern AI agents. Use 5–7 high-quality sources, prioritizing research papers and official technical documentation, and produce a visually polished executive brief that fits mostly within one screen: - A strong title and one-sentence conclusion - Three key findings - A compact comparison table for episodic, semantic, and working memory - One recommended architecture - Inline citations throughout Keep paragraphs short, avoid unnecessary preamble, and make the final response presentation-ready.

Analyzing · 4s

The user wants me to research how long-term memory is being implemented in modern AI agents, using 5–7 sources (research papers and official technical docs), and produce a concise executive brief with specific findings and a table. Let me start by searching for relevant sources.

Reading · long-term memory implementation AI agents 2024 2025 research paper · episodic semantic memory

Retrieval-Augmented Generation for Large Language Models: A Surveyarxiv.org
doiThe Power of Noise: Redefining Retrieval for RAG Systemsdoi.org
Securing LLM-Agent Long-Term Memory Against Poisoning: Non-Malleable, Origin-Bound Authority with Provenancearxiv.org
A Survey of Agent Memory in the Second Half: Towards Self-Evolving and Long-Horizon Agentsarxiv.org
doiAI Agents vs. Agentic AI: A Conceptual taxonomy, applications and challengesdoi.org
WAI -driven adaptive learning for sustainable educational transformationonlinelibrary.wiley.c…
Technical Report for Ego4D Long-Term Action Anticipation Challenge 2025arxiv.org
doiThe First Prompt-Native Semantic Operating System: EA-ARK-01 and the Architecture That Runs Inside Itdoi.org
From Human Memory to AI Memory: A Survey on Memory Mechanisms in the Era of LLMsarxiv.org
MemOS: An Operating System for Memory-Augmented Generation (MAG) in Large Language Modelsarxiv.org

TRYAsk anything…

Features

What Droplet does

Droplet uses the tools you already rely on to get work done.

Citations
4
1. Longformer: The Long-Document Transformer
  Beltagy et al. · arXiv
  1
2. Linformer: Self-Attention with Linear Complexity
  Wang et al. · OpenAlex
  2
3. Rethinking Attention with Performers
  Choromanski et al. · OpenAlex
  3
4. FlashAttention: Fast and Memory-Efficient Exact Attention
  Dao et al. · Google Scholar
  4
Droplet searches arXiv, OpenAlex, and Google Scholar, not only the open web.
Features
Research Retrieval
Ask a research question and Droplet goes to arXiv, OpenAlex and Google Scholar instead of the open web. Every claim in the answer is numbered and links to the paper it came from.
How do transformers handle long documents?2 sources
Sliding-window attention keeps the cost linear in the length of thedocument and low-rank projections get there another way.
OpenAlex · 10.48550/arXiv.2006.04768
Linformer: Self-Attention with Linear Complexity
Sinong Wang, Belinda Z. Li, Madian Khabsa, et al. · 2020
“…we demonstrate that the self-attention mechanism can be approximated by a low-rank matrix.”
Citations you can check
Each number in the answer opens a citation card with the paper, its authors and the passage it was cited for. Open the paper, its PDF or its BibTeX from there.
Scored and deduplicated
Results are scored on your machine before the answer is written, and overlapping sources are merged, so a paper found in two indexes is only cited once.
Source What Droplet uses
arXiv Abstracts, full-text PDFs and BibTeX
OpenAlex Citation counts, co-authors and DOI resolution
Scholar Related papers and peer-review status

Source	What Droplet uses
arXiv	Abstracts, full-text PDFs and BibTeX
OpenAlex	Citation counts, co-authors and DOI resolution
Scholar	Related papers and peer-review status

Press a shortcut and Droplet looks at what you’re looking at.

Features

Live Mode

Press U (CtrlShiftU on Windows) and drag over anything on your screen: an equation, a figure, an error message. Ask about it, and Droplet answers from what’s actually there.

en.wikipedia.org/wiki/Attention_(machine_learning)

Attention (machine learning)

Each token’s query is compared with every key. The scores are scaled, turned into weights with softmax, and used to mix the values:

Attention(Q, K, V) = softmaxQKᵀ√dₖV

where dₖ is the dimension of the keys. Without the scaling, large dot products push softmax into regions where its gradients are vanishingly small.

What does the √dₖ do here?2 sources

It scales the scores down, so softmax doesn’tsaturate when the keys are long and trainingkeeps its gradients. arXiv Wikipedia

ask about the selection…

Answers next to your work

The answer appears beside what you selected, with sources when Droplet checks them. Ask a follow-up from the same pill without capturing again.

Only when you ask

Capture is rate-limited on purpose. Live Mode runs when you press the shortcut, never as a screen recorder in the background.

Step	What happens
Capture	Press the shortcut, drag over the region and type your question
Thinking	Droplet reads the region and checks sources if it needs to
Answer	A reply beside your selection, with citations you can open

arxiv-filter

By Droplet

Scope arXiv searches to categories and date ranges.

scholar-graph

By Droplet

Follow citations through Semantic Scholar.

pdf-reader

By Droplet

Read local PDFs with page-level citations.

arxiv-filter

By Droplet

Scope arXiv searches to categories and date ranges.

scholar-graph

By Droplet

Follow citations through Semantic Scholar.

pdf-reader

By Droplet

Read local PDFs with page-level citations.

zotero-sync

Community

Push sources straight to your Zotero library.

sheets-mcp

Community

Log findings to a spreadsheet as you go.

github-search

By Droplet

Search code and issues across your repos.

zotero-sync

Community

Push sources straight to your Zotero library.

sheets-mcp

Community

Log findings to a spreadsheet as you go.

github-search

By Droplet

Search code and issues across your repos.

notion-export

Community

Save research notes to a Notion page.

wolfram-check

Community

Check the math with Wolfram|Alpha.

sandbox-run

By Droplet

Run code snippets in a sandboxed shell.

notion-export

Community

Save research notes to a Notion page.

wolfram-check

Community

Check the math with Wolfram|Alpha.

sandbox-run

By Droplet

Run code snippets in a sandboxed shell.

The tool system isn’t hardcoded.

Features

Systems & Plugins

Droplet’s tools aren’t hardcoded. Plugins add new ones, like scoping arXiv searches or saving sources to Zotero, without changing Droplet itself.

arxiv-filter

By Droplet

Scope arXiv searches to categories and date ranges.

scholar-graph

By Droplet

Follow citations through Semantic Scholar.

pdf-reader

By Droplet

Read local PDFs with page-level citations.

arxiv-filter

By Droplet

Scope arXiv searches to categories and date ranges.

scholar-graph

By Droplet

Follow citations through Semantic Scholar.

pdf-reader

By Droplet

Read local PDFs with page-level citations.

zotero-sync

Community

Push sources straight to your Zotero library.

sheets-mcp

Community

Log findings to a spreadsheet as you go.

github-search

By Droplet

Search code and issues across your repos.

zotero-sync

Community

Push sources straight to your Zotero library.

sheets-mcp

Community

Log findings to a spreadsheet as you go.

github-search

By Droplet

Search code and issues across your repos.

notion-export

Community

Save research notes to a Notion page.

wolfram-check

Community

Check the math with Wolfram|Alpha.

sandbox-run

By Droplet

Run code snippets in a sandboxed shell.

notion-export

Community

Save research notes to a Notion page.

wolfram-check

Community

Check the math with Wolfram|Alpha.

sandbox-run

By Droplet

Run code snippets in a sandboxed shell.

Loaded only when relevant

Plugins register their tools through a shared dispatch layer. A session loads only the guidance for the tools it uses, so more plugins don’t crowd it.

By Droplet, or the community

First-party plugins are marked By Droplet. Community plugins register the same way, through the same dispatch layer.

Plugin	What it does
arxiv-filter	Scope arXiv searches to categories and date ranges
pdf-reader	Read local PDFs with page-level citations
zotero-sync	Push sources straight to your Zotero library
github-search	Search code and issues across your repos

Learn more about Droplet

Built for real work.
Ready to install.

Install Droplet Get in touch

What’s next

Research Retrieval

Live Mode

Systems & Plugins

Built for real work. Ready to install.

Built for real work.
Ready to install.