Almanac is building an AI agent that actually understands your

PromptCube Intermediate 1h ago 177 views 10 likes 2 min read

Most AI agents feel like they have five-minute amnesia the moment you close the chat window. They can summarize a transcript or draft an email, but they don't actually know your business, your roadmap, or the subtle nuances of your team's workflow. I've been looking into how teams try to solve this, and it usually ends in a nightmare of manual data entry or broken OAuth connections.

The team behind Almanac (a YC S26 startup) just shared how they are tackling this specific friction point. Instead of treating memory as a secondary feature, they are building a "pre-compiled knowledge layer" that functions like a brain for your company.

The architecture of a proactive AI workflow

The core technical differentiator here is how they handle context. Most LLM agents are session-bound—they run a task, finish it, and the context vanishes. Almanac shifts this paradigm by using a dual-wiki structure to manage information:

  • Personal Wiki: This is strictly private. It learns your specific preferences, your personal schedule, and the people in your life. It's designed so the AI understands you without leaking your data to the rest of the team.
  • Company Wiki: This is the shared intelligence. It ingests data from connected sources to map out company goals, product roadmaps, and current blockers.

To make this work, they aren't just dumping text into a vector database. They are spending significant compute upfront on "pre-compiling" this knowledge. This allows the agent to move beyond reactive chat and into proactive task management.

Moving from reactive chat to proactive agents

If you've worked with LLM agents, you know the "prompt engineering" struggle: you have to constantly remind the AI who you are and what you are working on. Almanac attempts to automate this through two main features:

1. Proactive Notifications: Because the agent is constantly monitoring the background via a worker process, it can suggest actions before you ask. Imagine waking up to a notification saying, "I've drafted the fundraising pitch deck based on our last three meetings; want to review it?"
2. Long-Horizon Task Management: This is where real productivity happens. Most agents fail at tasks that require waiting. Almanac stays "on" even when you aren't. If it sends an email on your behalf and receives a reply four hours later, it doesn't just see a new message—it understands that message within the context of the ongoing project and drafts a relevant follow-up immediately.

Deployment and connectivity

Setting up these kinds of AI workflows from scratch is usually a massive headache involving custom API integrations and complex permissioning. Almanac aims to make this a beginner-friendly experience with one-click connections to:

  • Gmail & Calendar
  • PostHog
  • Granola
  • And various other workspace tools

The goal is to create an agent that feels less like a tool you use and more like a teammate who has been sitting in every meeting and reading every Slack thread. It’s a deep dive into what a true LLM agent needs to be successful in a real-world professional environment: persistent, contextual, and proactive.
Y CombinatorHermesAlmanac

All Replies (4)

Z
Zoe12 Novice 1h ago
A week isn't nearly enough time to see if the learning curve is actually worth it. If the AI needs time to gather data and get smart, a 7-day trial feels more like a teaser than a real test.
0 Reply
A
AlexHacker Expert 1h ago
How do they actually get around 2FA without getting flagged immediately? I've tried some similar automation scripts before, but the session persistence usually breaks the second a CAPTCHA pops up or the site detects a non-human pattern. It seems like a massive cat-and-mouse game.
0 Reply
C
ChrisPunk Novice 1h ago
If even a solo founder can't wrap their head around every single detail, how can any AI claim to have a complete grasp of a business? This feels like massive overpromising to me. It’s hard to buy into the hype when the math just doesn't add up.
0 Reply
S
SkylerDev Intermediate 1h ago
So basically, you're looking for a way to spend a fortune on a tool that just tells us our data is a mess? Sounds like a dream.
0 Reply

Write a Reply

Markdown supported