Your Memories + Agents = Command Center
No more time lost organizing. Instantly connect your sources, capture thoughts, and recall everything scattered. Constella links it all for you.
CONNECTED MEMORY SYSTEM
Ask across the contents of all your
files and apps
A pure magical experience.
YOUR SECOND BRAIN, IN MOTION
How Your AI Second Brain Works
Local files flow in, fuse into one connected mind, and become living insights — answered by your own context.
research-notes.md
q4-report.pdf
investor-email.eml
~/projects/notes
vaultYOUR ALWAYS-ON AGENT
A Central Dashboard for All
Your Assistants
Hand over your scattered to-dos. One agent works through them in the background — sourcing from your notes, tickets, and inbox.
Maximum Privacy
All 100% Stored On Your Computer.
Approved for Sensitive Work.
The maximum privacy possible. Pairs with your local models for HIPAA compliant work.
Tailored to your work
Answers that knowwhat you do all day
Constella reads in your sources, then speaks the language of your craft — with citations from your own work.
Your March lit review covers three of these directly. [1] reports a 14% effective-context drop past 32k on Mixtral-class models — but [2] contradicts that on Snowflake Arctic at the same depth. Your own notes from [3] flagged the eval methodology gap between them.
Three updates shipped: [1] rolled out per-seat billing on Tuesday. [2] from your roadmap.md is now blocked — the volume-tier idea assumes the old usage-meter, which we just removed. [3] from Linear is the unblocker; assign it before sprint planning.
Yes — [1] narrowed the scope of perpetual-confidentiality clauses in NorCal jurisdictions. That conflicts with section 4(b) of your current template [2]. Your memo from January [3] already drafted the carve-out language — it just needs the citation updated.
Their last B12 panel [1] showed 211 pg/mL — sub-threshold for metformin-associated deficiency. The Apr 9 visit note [2] mentions tingling in extremities, which fits. NICE guidance [3] recommends supplementing before adjusting dose.
Powerful Chrome Extension
Instantly clip, summarize, and recall as you read and write on the web.
Constella rides along in every tab — clipping what matters, summarizing as you read, and surfacing what you already know right as you draft.
Spectral-Tail Adapters: Protecting Principal Components in Parameter-Efficient Continual Learning
Parameter-efficient finetuning methods based on spectral decomposition have enabled progress in continual learning. In this paper we introduce Spectral-Tail, which utilizes the singular bases U and V of the pre-trained weights as a fixed reference frame to learn a low-rank update applied to the singular value matrix. A soft spectral penalty discourages updates aligned with the dominant singular directions, reducing interference while routing fine-grained adaptation into the long-tail coordinates.
1 Introduction
Large Language Models (LLMs) have achieved remarkable performance across diverse reasoning and generation tasks (Zhao et al., 2023; Minaee et al., 2024). However, adapting these models to new domains remains computationally expensive, as full fine-tuning requires updating billions of parameters.
Among PEFT approaches, Low-Rank Adaptation (LoRA) (Hu et al., 2021) has emerged as one of the most widely adopted. Motivated by the evidence that task-specific updates lie in a low-dimensional subspace (Li et al., 2018), LoRA freezes the pretrained weights and learns two trainable low-rank matrices.
Existing low-rank methods often suffer from interference between overlapping update directions, especially when models are adapted across sequential tasks. Since the largest singular values encode the most critical structure, modifications there disproportionately degrade prior knowledge.
To mitigate this, we propose a spectral regularization scheme that selectively penalizes updates to the dominant singular components while allowing greater flexibility in the lower-rank "tail". Our specific contributions are as follows:
- We introduce Spectral-Tail, a low-rank adaptation method operating over the singular values of a weight matrix, coupled with a soft regularization that steers updates toward the spectral tail.
- Different from existing continual PEFT methods (Das et al., 2026; Wang et al., 2023a), it requires no access to adapters from prior tasks, preserving the privacy of each user's task-specific data.
- We evaluate on a suite of continual learning tasks, matching state-of-the-art methods while increasing the stable rank of the weight matrix.
2 Related Work
Spectral LoRA variants. Leveraging the spectral properties of base weights W is a key strategy in PEFT. Many SVD-based approaches (Meng et al., 2024; Lingam et al., 2024) partition the spectrum to align trainable updates with the structure of pretrained matrices for more efficient tuning.
How I Use an AI Second Brain to Run My Business
Ever since I started saving everything into one place, meeting prep that used to take me an hour now takes five minutes, and the research that used to eat half a day takes twenty.
When you're running a business, most of the real work is hunting for context that's scattered across a dozen apps, old chats, and articles you swear you read last month. The fix isn't more notes; it's a system that recalls the right one at the right moment.
It could be a decision you made about this exact problem a quarter ago, and the reasoning behind it. Or the report you skimmed in February that's suddenly relevant to the call you're on today.
Select text and right-click to clip it
Sync your knowledgebases
Context Engineering in AI Brains
Personal knowledge tools promise perfect recall, yet most degrade into write-only archives. The bottleneck is rarely storage; it is context: surfacing the right memory at the exact moment of need.
Why retrieval is the hard part
Most retrieval systems treat memory as a flat store of chunks, but a real second brain has to weight recency, relevance, and the user's own
Instant Overlay
Instant Capture & SearchThroughout the Day
Hit ⌘⇧O while reading a paper, drafting in your inbox, or down a YouTube rabbit hole — capture the thought or recall what you already know, then disappear.
The product is labor.
Anthropic, which has yet to produce a single year of profit, commands a valuation in the same stratosphere. These numbers need an addressable market large enough to justify them.
There is only one market that big — the global market for human labor. The frontier labs are not selling software, they are selling labor itself, packaged as inference.
As we’re getting closer to that future, the bottleneck has shifted. The model is not the moat; distribution is. And distribution, increasingly, looks like in-person marketing work — pitching a different reality to people who already have the old one working fine.
The gentler interpretation is that the next decade of AI work looks less like coding and more like sales.