🧠 Your agent has amnesia

Be honest. How many times this week did you re-explain your project to an AI agent? The folder structure, the thing you decided yesterday, the bug you already fixed?

We lost count. And it's getting more painful, because agents are no longer one-off helpers. This week Perplexity launched always-on "Teammates" agents, and Meta bought Conductor, a startup that lets developers run several coding agents in parallel.

More agents, running longer, on more of our work. Which raises a simple question: what happens to everything they learn along the way?

🔭 The big picture: memory is the next layer

In Issue #1 we drew the agent stack: backend, interface, memory, agents, security and end-user apps. Memory was the quiet one. Not anymore.

  • Forgetting is expensive. Every time an agent starts from zero, you pay for it twice: once in your time re-explaining, and again in tokens as it re-reads everything.

  • Investors noticed. Memory startup Engram raised $98M back in June with a pitch built around cutting token costs.

  • Developers want memory they own. As agents touch more sensitive work, "where does my agent's memory live?" is becoming a real question.

Our read: the agents that win won't just be the smartest. They'll be the ones that remember.

🚀 Startup Spotlight: CMEM

The company to look out for in agent memory

CMEM started with a delightfully simple idea: while one AI works, a second AI takes notes.

That idea became claude-mem, an open-source memory engine created by Alex Newman. You install it, keep working, and it quietly records what your agent did, what you decided, and why. Next session, your agent picks up where it left off instead of asking you to start over.

A few things we love about it:

  • It's local-first. Your memory lives on your own machine by default. Cloud sync is opt-in.

  • It works across tools. With CMEM Pro, one private link lets Cursor, Codex CLI, Gemini CLI and other agents all share the same memory.

  • It saves money, not just time. Its code-exploration feature says it's 6–12× cheaper than having an agent read files in full.

  • It's built for teams too. A shared memory means a new teammate can ask the team's "brain" instead of interrupting everyone.

The traction: developers clearly feel the pain. CMEM's own numbers from June showed more than 94,000 GitHub stars and 6.6 million AI sessions compressed in a single day.

Why it matters: if memory becomes the layer that makes agents useful, whoever builds the most trusted version of it matters a lot. CMEM is betting on open source and user ownership, and so far developers are betting with them.

Want to try it? Run npx claude-mem@latest install

For investors: curious about the companies in our portfolio? Drop a note to [email protected] and we'll share more. We'd love to hear from you.

🔬 Research corner: AI that edits AI's brain

This one surprised us!

On November 2, MIT CSAIL researchers published "Sparks of Weight-Space Programming". In short, a frontier AI model can now inspect another neural network's raw weights , the millions of numbers that make up its "brain" and understand, fix, or even write them, without a training run.

What they showed:

  • It can read. Given only the weights of a small game-playing network, the model identified which game it had been trained to play.

  • It can write. It created small game-playing networks from scratch that outperformed networks trained the usual way for 100 million steps.

  • It can fix. Asked to improve a small language model's math ability, it identified the cause of reasoning loops and edited just 0.016% of the model's weights. Its math score rose from 7.9% to 48.1%.

We also appreciated the researchers' candor about the limits. The model couldn't identify a people-pleasing model from its weights alone, nor improve a model that didn't have the looping problem. Their takeaway: weight editing can remove an obstacle hiding an existing ability, but it can't create an ability that isn't already there.

Why should founders care? Picture checking a model for hidden problems without running a single test, or fixing a bad behavior in minutes instead of retraining for weeks. It's early, but if this holds up, the cost of improving models could drop dramatically. The team is building a benchmark next, and we'll be watching.

📡 Signal vs. noise

A few more things that caught our eye this week:

  • Sentra raised a $150M strategic round led by Capital One Ventures to protect the data AI agents touch. Every new agent is a new way for data to leak, and buyers know it.

  • Sierra acquired Anon, whose tech lets agents sign in and fill out forms on websites that have no API. Translation: agents are learning to use the web the way we do.

  • Fal raised a $1.3B Series F at a $16B valuation, led by Bessemer. Infrastructure for running AI models is still one of the hottest places to be.

  • Nooks announced $500M in growth capital alongside a platform where sales reps and AI agents share the work. Sales teams are becoming human-plus-agent teams.

🔢 One number: $500M

That's roughly what Palo Alto Networks is reportedly in talks to pay for Clutch Security, a startup that tracks and secures the credentials AI agents use. Remember "who is this agent, and what can it touch?" from Issue #1? That question now has a price tag.

🤔 Contrarian corner: memory is the new lock-in

Here's a slightly uncomfortable thought. For years, the stickiest thing a software company could own was your data. In the agent era, it might be your agent's memory.

If one platform holds everything your agent has learned about your work, switching tools means starting from zero. That's great for the platform and not so great for you.

Our take: developers will increasingly choose memory they can take with them, which is exactly why open, portable approaches matter. Before you pick an agent tool, ask a simple question: if I leave, does my agent's memory come with me?

🎤 Super Demo Day is 10 days away!

This is it: 2026 Super Demo Day lands in San Francisco on Monday, November 16. Twenty startups, four demo sets, and a room full of people who build things for a living. It's presented by Open Source for AI, in partnership with Gravitational Ventures and AWS.

If you've been meaning to apply, this is your nudge.

💬 Your Turn

For investors: curious about the companies in our portfolio? Drop a note to [email protected] and we'll share more. We'd love to hear from you.

Know someone who'd enjoy this? Send it their way.

Keep Reading