Why context and a knowledge base decide it, and how I keep mine
What my AI starts with is its context, and the context is the part I can improve. The model I open tomorrow won't have learned anything I corrected today.
So that's where I work on it. In personal AI usage, 3 AI agents share one context: Hermes, Codex and Claude Code, across different servers and laptops.
Why context matters
When an agent starts, the model behind it knows what it learned in training and nothing about my work. My rules and yesterday's decisions reach it only if they're in front of it, in the files it loads and what it retrieves.
Context is how the AI gets closer to me. I think the more it holds about what I know, how I work and what we've done together, the better its answers fit me.
I set things up this way because I kept switching. A different agent for a different job, a new session, another machine. Each switch left some context behind.
And a fix made in a chat lasts only if it lands somewhere the next agent will read. When the conversation holds the only copy, the next agent starts over.
How I do it: 3 places for context
Mem0 holds my interactions. It runs alongside an agent session and saves details as the work happens, so a later session can find them.
GBrain holds what I know. It keeps pages, facts, links and timelines about the people and projects I work with, and an agent searches it or adds to it on purpose.
Both are services I run myself, and they work across all my projects and from any of my machines. I first wrote about them in After the laptop closes, and I call the pair my knowledge base. Andrej Karpathy describes a close idea in his LLM wiki: a personal knowledge base the AI builds and keeps current.
The third place holds how I work, per project. Each project's repository carries Markdown files that agents read when they work there. In my writing project, those are the working rules, my voice guide, a review checklist and short procedures for jobs that come back.
Git carries those files to every device. But the files in a project's repository belong to that project alone.
Keeping the context outside the agents is what keeps my options open. When a better agent or model appears, I can switch without starting from scratch, because once it's connected it picks up the same context. I argued for that kind of portability in The Open-Source Lesson for AI, and this is how it works in my own setup.
That's how an AI can get better while the model stays the same: dozens of small adjustments that live in the context and move between sessions and agents. A rule I write down once is in front of the next agent that works in that project, on any machine.
What it costs
The closer the AI gets to me, the more my knowledge base holds about me. In AI made the desktop relevant again I wrote about deciding what leaves my machine, and this is the context I most want to keep on my side.
A wrong fact about me can also reach every agent, so I have to think about permissions and prompt injection at the level of the store. OWASP lists memory and context poisoning among its top 10 risks for AI agents.
You need to be sure the setup is replicated properly across agents and machines. If it isn't, you find out slowly, through the results of the work.
Retrieval takes time.
The Markdown files cost something too. If I start on one machine before syncing what I wrote on the other, the agent reads yesterday's rules. Syncing first is now the opening rule of the project.
And a knowledge base needs a steward. Automatic capture repeats itself and some of what it holds goes out of date.
Mem0 and GBrain now each have a "dream" step for that. GBrain's runs in the background on my server, linking pages, pulling out facts and checking for what has drifted; Mem0's merges duplicates and prunes stale memories, then shows me the changes before applying them.
The AI does the bookkeeping. Deciding what's right is my job.
What's next
For a team or a company, I think the knowledge base becomes a key part of how its AI works. If I were choosing agent tools for a team, I'd decide early where it lives, because whatever a tool holds only inside itself leaves with it.
It needs the right access levels, for the people inside and for partners outside. And it needs levels of its own.
It's much the same question companies already face with data, now for a new tool. There's my knowledge, my team's and my company's, plus what each process needs. And there's my way of working, my team's way and the company's way.
Skills, and the stores that share them, are one way to automate the last part. They package the Markdown files and configuration that agents load, so each level can write its way of working down once.
For the knowledge, the connection matters as much as the store. I'd choose memory providers that agents reach through MCP, the open protocol that Codex, Claude Code and Hermes all support, so a new agent plugs into the same knowledge base without a custom integration.
In What an AI experiment reveals about our organizations I wrote that keeping company context current is work the organization has to own. Deciding where that context lives, and who sees each level, is part of the same job.
If your team changed its AI tool tomorrow, what would it forget?