https://x.com/i/article/2112570453860466688
On Building Persistent Agents
More and more, I write about the importance of owning your intelligence stack.
That includes owning your agent's memory. Beyond the harness and models, memory is a key part of the intelligence stack you want to manage well to unlock agents' frontier capabilities.
How does this look in practice and why does it matter?
Let's focus on persistent agents, which is one of the most exciting new applications today.
Persistence has become a core capability of recent models like Opus 5.5, GPT -6 Astra, and Gemini 4 Argon.
Recently, four Gemini 4 Argon agents beat Minecraft in 27 hours. Along the way, they recovered items, revised strategies, and shared lessons through a shared memory.
Models are getting better at choosing what to persist, and how and when to use it. What they persist needs a place that outlives any one session, and you should fully control that.
Recently, I have been building and testing my own personal agents. What I have found is that memory is key to what makes them useful.
Specifically, I want personal agents for long-running tasks.
In the standard setup, memory usually lives inside one harness on one machine, so when I switch to a different agent, it can't see what the first one learned.
I have been exploring a few solutions.
Recently, I have been testing @FulcraDynamics, which gives your agents a shared context layer you own.
One agent writes to it, and the next one picks up from there, across tools and sessions.
So I tested it on a real handoff between two coding agents.
I gave Claude Code a small coding task and had it save its progress, decisions, and next steps to a shared Fulcra workspace after the first step. Then I opened a fresh Codex session and only told it the workspace name.
Codex read the workspace, figured out what was done and what was left, and finished the last two steps without redoing the first. All the tests passed, and I never pasted anything between the two agents.
This is extremely useful for long-running agents.
Long-horizon agents can run for days, get restarted, switch models, and run out of context. When their state lives in Fulcra, a fresh session reads where things stand and picks up the work from there. This lets work persist, which you need to build agents for complex, long-running tasks.
You can also decide what each agent sees. You can share one folder with a person or their agent and keep everything else private.
It's your context, and you can export or delete it at any time.
You can start by adding their agent skills with npx skills add fulcradynamics/agent-skills
Try Fulcra here: https://www.fulcradynamics.com/product/start-for-free?utm_source=x&utm_medium=paid_social&utm_campaign=omarsar0
Thanks to Fulcra for partnering on this post.