Forgetful Assistants Learn to Remember: Hermes Meets Its Vault

Repeating your own story at the start of every chat wears you down. You open the agent and it has forgotten your goals; it knows nothing about last month, so you type the same sentences again. The video uses a harsh image: the memory of a goldfish in a bowl. As the Weaviate team puts it in their context engineering essay, the window is a scarce resource, and every step needs a decision about what goes inside. Without that decision the agent starts guessing.

The side-by-side test in the video lands hard. Asked what we did last month, the setup with memory answers fast and in detail, while the same question in the setup without memory spins for two and a half minutes and ends thin. The MongoDB guide draws the line I keep returning to: short-term grip and long-term memory are different things. Missing the second one, the agent cannot tell where to look or what to search for.

The fix goes by the name infinite context engine in the video. The idea is plain: Hermes is the agent doing the work, Obsidian is the vault holding the facts, and you wire the two together into one shared brain. What won me over is how concrete the payoff sounds: a more personal tone, sharper answers, speed because it pulls from memory, and no more daily re-introduction. The Hermes side checks out in the NousResearch docs: an agent with lasting memory under a public license, learning your projects after setup and keeping how it solved each problem.

Open the vault and two things stand out: everything is linked, tagged and color-sorted, while the notes run deep about who you are and who is on your team. That goes far beyond the daily notes Hermes keeps on its own. That is why this question works: looking at my recent work, what should I automate this week to reach more people? The answer stops being a template and turns into a plan built for you. The Obsidian side is equally clear: per the obsidian.md site, it is a free and flexible app for your private thoughts, with everything sitting in plain files that any agent can read.

We all know the old path: a blank page opens, you paste your goals, paste your client notes, explain who you are, and thirty seconds go to waste in every chat. Worse, the agent never hears about work done in another tool, so notes slip away, context slips away, answers go flat. The video says the same about ChatGPT: it knows nothing of your life outside it. The new path is a one-time job: point the agent at the vault a single time, and every task starts by reading real notes, even pulling a line you wrote months ago. Wikipedia traces this power to links and the graph view in its Obsidian article: as notes connect, the vault turns into a map.

Now the five layers. Layer one is the vault: your notes become memory. The video uses a small recorder called Omi, an open project that catches conversations, turns them into text and remembers them, with two thousand memories exported into the vault. The Omi docs say it plainly: it captures your conversations, turns them into text and remembers them. The whole project sits on GitHub in the open, and you can run it on your own box. Linking it to Hermes takes one sentence: I have Obsidian on my computer, use it for memory from now on.

Layer two is the link: Hermes gets to read everything inside the vault. Layer three is the common mind: Hermes, Claude and OpenClaw all feed from the same vault and speak about the same work in a single voice. Layer four is recall: it finds last month notes and lays them out. Layer five is the loop: Hermes writes the new notes, so the longer the use, the sharper it gets, with zero manual effort. As Weaviate puts it, fetching selection from outside storage is the secret of a dependable agent: not stuffing everything into the window, but picking the right piece at each step.

The trick saved for last: do not run Hermes on its own, place it inside one mission board. That is what the video calls its Agent OS: everything visible in one place, no hopping between tabs, clear what got built, and the vault stays fresh because every job feeds the brain. The flow runs like this: you hand Hermes a job, a note lands in the vault, Claude reads the same note next morning. The natural-language scheduling in the NousResearch docs fits the same board: reports and briefings run while you are away.

The video hands out two starter commands and I wrote both down. One: looking at my recent work, what could I automate today to save time? Two: the board-building command, asking for a board that tracks videos, contacts and daily tasks while reading from the vault. Trying the second myself, I would narrow it to one purpose instead of a general board; a focused board pays back sooner. The MongoDB warning belongs here: without long-term memory the agent falls back to guessing, which is why the command opens with look at my past.

A closing thought: early movers win here, because most people still paste the same facts into a blank chat, and that window will not stay open forever. The creator of the video offers the full template, the ready commands and a thirty-day roadmap inside his own community, while a large free group shares notes on the side with membership past ninety-three thousand. My stance is simple: take the template, start with your own vault, and the rest compounds as you keep using it.