How to Give Your AI Agents Persistent Memory with an Agent Operating System

Usage & License

Use the prompts and downloadable templates on this page in your own personal and commercial projects. What's not allowed: reselling, repackaging, redistributing, or republishing this material — on YouTube, Gumroad, paid courses, newsletters, or anywhere else. Send people to komputermechanic.com instead. Violations get reported and taken down.

We are building the memory layer for my Hermes agent crew: a wiki at ~/wiki of plain
markdown the agents read before answering and write to after working. Chat history gets
summarised away; this does not.

  processed/   what agents WRITE - my world (personal/, business/), outside
               material (research/)
  raw/         the originals, verbatim, never edited - so any claim can be
               checked against its source

Build order: backend first - wiki, skill, habits, ingestion, feeds and the night
jobs - until the whole thing works and is proven. Then the dashboard, so I can see
what the agents know. Git last, as the safety net. I feed you numbered prompts one
at a time.

Read these two first (if unreachable, say so and carry on):
1. https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f - the idea: the
   model maintains a persistent wiki instead of retrieving at query time.
2. https://hermes-agent.nousresearch.com/docs - profiles, skills, SOUL.md, cron.

YOU ARE A BUILD TOOL, not a crew member. Files, scripts, routes, jobs: you. Habits the
agents must own: I paste those into the agents myself. A scheduled job that thinks is a
Hermes cron agent turn on a named profile, never a shell script imitating one.

STANDING RULES: one prompt at a time, never run ahead. Back up before changing.
Absolute paths in anything scheduled. Verify by looking, never by claiming. When you
finish a prompt, say what you did and what you are waiting for.

BEARINGS: report whether Hermes runs directly or in a container here. If
containerised, ~/wiki must be the SAME real path for every profile and for you - pin
CREW_WIKI_DIR in the container env and symlink each profile's home/wiki to it,
otherwise per-profile HOMEs give every agent a different empty wiki. Also report:
every profile and its HERMES_HOME, what is at ~/wiki, whether systemd --user and git
are available, and date +%Z (a hint - the scheduler's zone is verified later).

Close with your understanding: what we are building and why it beats chat history, the
two layers, the build order, your role versus the agents'. If anything here seems
wrong for THIS machine, say so now. Build nothing yet.

We are building the memory layer for my Hermes agent crew: a wiki at ~/wiki of plain
markdown the agents read before answering and write to after working. Chat history gets
summarised away; this does not.

  processed/   what agents WRITE - my world (personal/, business/), outside
               material (research/)
  raw/         the originals, verbatim, never edited - so any claim can be
               checked against its source

Build order: backend first - wiki, skill, habits, ingestion, feeds and the night
jobs - until the whole thing works and is proven. Then the dashboard, so I can see
what the agents know. Git last, as the safety net. I feed you numbered prompts one
at a time.

Read these two first (if unreachable, say so and carry on):
1. https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f - the idea: the
   model maintains a persistent wiki instead of retrieving at query time.
2. https://hermes-agent.nousresearch.com/docs - profiles, skills, SOUL.md, cron.

YOU ARE PART OF THE CREW, and for this build you also wear the builder's hat: you
construct the files, scripts, skills and jobs the whole crew will run on - yourself
included. The discipline that comes with the hat: rules and habits are never saved
from memory - they are written into identity files VERBATIM with your file tools,
your own SOUL.md included, when the prompt for that comes. Wherever a later prompt
says Claude Code, it means you. A scheduled job that thinks is a Hermes cron agent
turn on a named profile, never a shell script imitating one.

STANDING RULES: one prompt at a time, never run ahead. Back up before changing.
Absolute paths in EVERYTHING - your own home folder may not be the wiki's home (in
containers each profile can have its own), so pin the wiki's real path once and use
it everywhere. Verify by looking, never by claiming. When you finish a prompt, say
what you did and what you are waiting for. If your session resets mid-build, I will
re-paste this briefing and tell you which step we are on.

BEARINGS: first report which profile you are running as, and prove your terminal and
file tools work. Then: does Hermes run directly or in a container here; every
profile and its HERMES_HOME; what is at ~/wiki as the DEFAULT profile sees it, not
just as you see it; whether systemd --user and git are available; date +%Z (a hint -
the scheduler's zone is verified later). If containerised, ~/wiki must be the SAME
real path for every profile - pin CREW_WIKI_DIR in the env and symlink each
profile's home/wiki to it.

Close with your understanding: what we are building and why it beats chat history,
the two layers, the build order, and the builder's hat you wear on top of your crew
role. If anything here seems wrong for THIS machine, say so now. Build nothing yet.

Build ~/wiki. If it exists and is not empty, STOP and tell me what is in it.

  ~/wiki/
  |-- processed/          what the agents WRITE
  |   |-- personal/       me, my household, anything not the business
  |   |-- business/       what we sell, who buys, who we pay
  |   \-- research/       what we read - their claims, never our facts
  |-- raw/                originals, verbatim, never edited
  |-- index.md            the catalog: one line per page, each a link to that page,
  |                       a heading per parent, Research its own section - agents
  |                       read this FIRST
  |-- log.md              append-only diary, one line per change
  \-- WIKI.md             the rulebook

personal/, business/ and research/ in these prompts always mean the folders inside
processed/. Do not create sub-folders yet; they come later, when real material needs
them.

WIKI.md rules - write them into the rulebook, short and clear:

- Where a page goes: what I make or sell goes in business/; what we read to learn
  goes in research/; everything else about me goes in personal/. A person goes where
  the money flows - a customer, a supplier and a partner are different pages. When
  two folders both fit, file the page where I would look for it later (a
  competitor's price list goes in research/ - I open it to check the market). Every
  page has exactly one home.
- Names and links: page names are lowercase-with-dashes and unique across all
  folders. The first mention of anything that has a page becomes a [[link]] - but
  never inside emails, handles, headings or code, and never a first name or a
  common word that could point at the wrong page.
- Writing: pages are short and factual, a one-line summary on top, a date on
  anything that can go stale. A repeated fact gets a counter, never a second copy.
  Contradictions are flagged side by side, never overwritten. Known aliases are
  listed on the page, and the page's index line ends with (aka ...).
- Bookkeeping: every edit also updates index.md and log.md. Define the exact line
  formats in the rulebook - the nightly checks will hold everything to them.
  Underscore paths are machinery: never counted, never indexed.
- Never: small talk, and never secrets - a credential, code or bank detail is
  recorded by name only, never by value.
- WORKFLOWS, both habits in full (this only stores them in WIKI.md - the agents get
  them in a later prompt): RECALL - search the wiki before answering anything it
  might know; answer from it when it knows, say "Not on file" when it does not.
  FILE - after real work, file the lasting part through wiki file, never by hand.
  Leave ingest and fact-vs-inference as empty headings for later.

Show me the tree and the rulebook.

Subscribe to unlock the full prompt set

You've just walked through the first 3 prompts. Drop your email below to get the rest. Orchestration, edge cases, polish, plus the dashboard template. 100% free for subscribers.