In this guide to the Hermes memory budget, here’s what we’re going to cover: the two costs people conflate, the leak memory fixes and the one it can cause, and the cache pricing that makes the same vault cheap behind one brain and expensive behind another.
Short answer
- Memory on disk is free; memory in context is paid on every reread — that’s the budget.
- A vault kills the re-explaining leak; a bloated vault becomes a rereading leak.
- Keep instruction files on a diet — Anthropic cut theirs ~80% with the same quality.
- Run reread-heavy work on cache-cheap brains and cap anything unattended.
The Hermes memory budget problem, stated plainly
Memory has two costs people conflate. There’s the cost of having it — markdown on disk, effectively free — and the cost of reading it, which happens inside the context window on every session, every loop tick, every sub-agent call.
An agent rereads its working context at every step, like a worker opening the full job folder before every action. So whatever memory you load per session gets paid for dozens or hundreds of times per task. That’s the budget to manage — not the disk.
The leak memory fixes vs the leak memory causes
First, credit where due: re-explaining who you are at the start of every session is one of the sneakiest token leaks going, and a memory vault kills it. Explaining once beats explaining hundreds of times a year.
But a bloated vault swings the other way. Load everything into every session and you’ve replaced the re-explaining leak with a rereading leak. The goal is a distilled vault: facts, decisions and voice — not raw dumps.
| Habit | Effect on the memory budget |
|---|---|
| Distil notes; prune stale plans | Small context load, paid many times — cheap |
| Dump transcripts and logs raw | Huge load, paid many times — expensive |
| Keep persona/instruction files lean | Anthropic’s own team cut theirs ~80% with the same quality out |
| Scout pattern for big digs | A sub-agent burns its own context and reports back a summary — your session pays for the conclusion, not the search |
🔥 Want this set up without the guesswork? The token-minimisation playbook covers exactly this — memory that makes sessions smarter without making them heavier. Inside the AI Profit Boardroom you get the Agent OS as a ready-to-install file, a 30-day roadmap, daily tutorials the same day new tools ship, and four live coaching calls a week where you share your screen and get unstuck. → Get access here
Cache pricing: where the memory budget gets cheap
Because agents reread the same memory constantly, those rereads are cache hits — and cache pricing varies wildly by model. Cache reads on DeepSeek V4 Pro cost roughly 276 times less than Fable 5, with a ~92% hit rate on OpenRouter.
Practical upshot: the same vault costs pennies behind a cache-friendly brain and real money behind a frontier one. Load memory generously on the volume profile, sparingly on the frontier profile you reserve for planning. The routing logic is in the DeepSeek pairing guide.
A working Hermes memory budget, in rules
- Seed and keep a distilled vault — ten sharp notes beat a hundred flabby ones.
- Prune weekly with
/journey; stale plans cost tokens forever. - Keep SOUL and instruction files on a diet — every rule rides along with every session.
- Send big reads to a scout sub-agent; keep the summary, drop the search.
- Match memory-heavy work to cache-cheap models.
- Cap anything unattended — loops and cron pay the memory budget while you sleep.
Setup first if you haven’t: the memory setup guide; capacity questions live in the memory limit guide.
Quick budget wins this week
- Prune five stale notes — each one was riding along with every relevant session.
- Halve your longest instruction file; check nothing breaks. It usually doesn’t.
- Move one big recurring dig to a scout sub-agent and keep only its conclusions.
- Check the cache hit rate on your volume profile — a miss quietly multiplies the whole bill.
FAQ
What is the Hermes memory budget about?
Not disk — context. Whatever memory a session loads gets reread at every step, so the budget is the size of what you load times how often agents reread it.
Does a memory vault save tokens or cost tokens?
Both, depending on discipline. It kills the re-explaining leak, but a bloated vault becomes a rereading leak. Distilled notes are the difference.
How big should the vault be?
On disk, as big as you like. Per session, small: the facts, decisions and voice relevant to the job — not transcripts.
What’s the scout pattern?
A sub-agent digs through the big messy stuff in its own context and reports a conclusion, so your main session pays for the summary rather than the search.
Which models make memory cheap?
Cache-friendly ones — DeepSeek V4 Pro’s cache reads run ~276x cheaper than Fable 5, which is exactly the pricing that suits constant rereads.
Do unattended agents affect the budget?
More than anything else — every loop tick reloads context. Set the caps before you sleep, not after the bill.
The bottom line on the Hermes memory budget
The Hermes memory budget is a context problem wearing a storage costume: distil the vault, diet the instruction files, scout the big reads, run rereads on cache-cheap brains, and cap the unattended stuff. Memory should make sessions smarter and lighter — if yours got heavier, it’s a pruning week.
About Julian Goldie
I run Goldie Agency, a 7-figure SEO agency, and teach this daily on a 400K+ subscriber YouTube channel. 240+ client projects on Upwork at a 100% job-success score, 10+ years through every major Google update. My systems are in the AI Profit Boardroom; my link building book is free here.
Related reading
Last updated August 2026. This is the living guide to Hermes memory budget — it gets updated as the tools change.