Current structure for my VPS and summarized below
This has helped immensely with a variety of projects. Would explore it with either codex or claude code. Then install either on your VPS. I have both. Which can work on your phone, desktop, laptop etc. and means you have access to the same virtual machine wherever. Every project is now connected to my overall knowledge source that constantly grows and improves. With research that can be done overnight if you set it up correctly. This is the ingest process. It also include my AI council which I've posted about before as I have my openclaw also on it that is connected to Codex.
Sky's the limit
I've been using
@Vultr . Will include a link that will will get you $250 in credits
vultr.com/?ref=9905573-9J
AI VPS Knowledge System Blueprint
This VPS is set up so AI coding agents do not rely on chat history as memory. Instead, they use a Git-backed knowledge wiki plus small agent-specific memory files.
Core Idea
There are two memory layers:
1. Agent instructions and session memory
- Global behavior rules for each AI agent.
- A current handoff file for “resume here” context.
- A small persistent memory index for durable gotchas.
2. Knowledge wiki
- A private Git repo that stores project facts, hypotheses, rules, decisions, raw research, changelogs, and wiki articles.
- This is the source of truth.
Recommended Folder Layout
~/knowledge/
CLAUDE.md # operating manual for agents using the wiki
INDEX.md # list of domains/projects
CHANGELOG.md # append-only activity log
decisions/ # dated decision journal
quality/ # quality gates/checklists
syntheses/ # larger cross-topic analyses
tools/ # docs for local tools/scripts
meta/
sources-of-truth.md # which file wins when facts conflict
project-a/
knowledge.md # confirmed durable facts
hypotheses.md # unconfirmed patterns to test
rules.md # proven default rules
articles/ # wiki-style explanations
raw/ # immutable source material
project-b/
knowledge.md
hypotheses.md
rules.md
articles/
raw/
Agent Startup Routine
At the start of meaningful work, the agent reads:
~/knowledge/CLAUDE.md
~/knowledge/INDEX.md
~/knowledge/CHANGELOG.md
~/knowledge/{project}/rules.md
~/knowledge/{project}/knowledge.md
~/knowledge/{project}/hypotheses.md
If resuming prior work, it also reads:
~/.agent/HANDOFF.md
Knowledge Rules
Use these rules to prevent drift:
- raw/ is the immutable source layer.
- knowledge.md stores confirmed facts.
- hypotheses.md stores plausible but unconfirmed patterns.
- rules.md stores things agents should do by default.
- decisions/ stores choices that affect future work.
- CHANGELOG.md records what happened.
- Existing knowledge files should be append-only unless a specific section is being updated.
- If two files conflict, the source-of-truth map decides which one wins.
Ingest Flow
External info should not go straight into rules. It should move through this pipeline:
source document / research / transcript
-> project/raw/
-> compile/extract
-> knowledge.md or hypotheses.md
-> rules.md only after repeated confirmation
Sync Flow
The wiki should be a private Git repo:
~/knowledge/
-> commit
-> push to private GitHub/GitLab repo
-> other machines pull from that repo
Optional compatibility mirror:
canonical knowledge repo
-> mirror repo
-> desktop/laptop/sandbox pulls mirror
Automation To Add
Useful scripts/hooks:
session-end hook # auto-commit and push knowledge changes
daily compile cron # compile new raw files
sync health check # alerts if pushes or pulls are stale
backup script # encrypted offsite backups
secret guard # blocks agents from reading credentials
Security Rules
Never put these in the knowledge repo:
.env files
SSH keys
API keys
OAuth tokens
cloud credentials
database passwords
backup passphrases
private auth JSON files
Use placeholders in docs:
<server-user>
<your-domain>
<github-user>
<project-name>
<service-name>
<backup-bucket>
<secret-store>
Shareable Diagram
flowchart TD
User[User] --> Agent[AI Agent Session]
subgraph AgentMemory[Agent Memory Layer]
GlobalRules[Global Instructions]
Handoff[Current Handoff]
SmallMemory[Persistent Agent Memory]
end
Agent --> GlobalRules
Agent --> Handoff
Agent --> SmallMemory
subgraph KnowledgeWiki[Private Git-Backed Knowledge Wiki]
Manual[CLAUDE.md / Operating Manual]
Index[INDEX.md]
Changelog[CHANGELOG.md]
Rules[project/rules.md]
Facts[project/knowledge.md]
Hypotheses[project/hypotheses.md]
Articles[project/articles/]
Raw[project/raw/]
Decisions[decisions/]
Quality[quality/]
end
Agent --> Manual
Agent --> Index
Agent --> Changelog
Agent --> Rules
Agent --> Facts
Agent --> Hypotheses
Agent --> Quality
Agent --> Work[Code / Research / Operations Work]
Work --> NewLearning[New Facts, Decisions, or Patterns]
NewLearning --> Changelog
NewLearning --> Facts
NewLearning --> Hypotheses
NewLearning --> Articles
NewLearning --> Decisions
subgraph IngestPipeline[Ingest Pipeline]
Source[External Source]
Ingest[Ingest Script]
Compile[Compile / Extract Script]
end
Source --> Ingest
Ingest --> Raw
Raw --> Compile
Compile --> Facts
Compile --> Hypotheses
subgraph Automation[Automation Layer]
StopHook[Session-End Hook]
AutoCompile[Daily Auto-Compile]
GitPush[Verified Git Push]
Backups[Encrypted Offsite Backup]
HealthChecks[Sync / Cron Health Checks]
end
Agent --> StopHook
StopHook --> GitPush
AutoCompile --> Compile
KnowledgeWiki --> GitPush
KnowledgeWiki --> Backups
GitPush --> RemoteRepo[(Private Git Repo)]
subgraph OtherMachines[Other Machines / Sandboxed Agents]
Laptop[Laptop / Desktop Clone]
Sandbox[Sandboxed Agent]
SyncScript[Sync Script]
IngestEndpoint[Private Ingest Endpoint]
end
Sandbox --> SyncScript
SyncScript --> IngestEndpoint
IngestEndpoint --> Raw
Short Version
The system is a private wiki for AI agents. Agents read it before work, update it after work, file raw sources into
raw/, compile them into facts or hypotheses, and promote only proven patterns into rules. Git, hooks, cron jobs, and backups keep the whole thing synced and durable