One of the first things I noticed about Hermes is that it forgets.
Context compression gradually removes older conversations, while every new session starts with a clean slate. The agent that learned your preferences on Monday may have no idea what you told it on Wednesday.
I solved this with two layers: memory and a personal wiki.
Hermes’ built-in memory is a small text file injected into every conversation. I use it for constants: my timezone, communication preferences, working style, and other details the agent should always know.
It handles the basics well, but it has room for only about 2,000 characters. That makes it unsuitable for everything Hermes learns over time.
That’s where the wiki comes in.
I created a simple directory with three categories:
- Entities for people, companies, tools, and other things
- Concepts for ideas and recurring themes
- Comparisons for options, tradeoffs, and decisions
Every page includes front matter and links to related pages. When Hermes conducts research, durable facts go into the wiki. When a recurring task produces useful intelligence, that goes in too. The same applies when I discover a new pattern, preference, or insight worth preserving.
Over time, the wiki becomes more valuable. After a few weeks, Hermes knows things about my job market, reading habits, and tool preferences that would otherwise take hours to reconstruct.
The key is keeping the two layers separate:
Memory holds constants. The wiki holds accumulated knowledge.
A timezone belongs in memory. Research about a company belongs in the wiki. A communication preference belongs in memory. A comparison that informed a decision belongs in the wiki.
Once Hermes understands that distinction, it knows where new information should go.
If you self-host Hermes, set up both layers from day one. The value may seem small at first, but the knowledge compounds with every session. After enough time, you are no longer working with an agent that repeatedly starts over. You are working with one that builds on what it has already learned.