System prompt
Always loadedSterling’s identity, safety rules, tools, persistent memory and skill catalog. Stable and always present.
Sterling 2.0 operating architecture
How the agent system thinks, remembers and scales
This reference explains the working context, persistent memory, skills, compaction, durable evidence, local capacity, Cloudflare execution, and the signals that justify another Hermes agent or provider account.
Core operating principleAgents reason.Deterministic workers execute.Humans approve consequential actions.Durable systems retain truth and proof.
Plain-English orientation: The context window is the temporary whiteboard. Persistent Memory is the small set of facts placed on every new whiteboard. Skills and databases are the organized library we retrieve from only when needed.
01
The complete architecture
Different kinds of knowledge live in different layers. Efficiency comes from loading only the layer required for the current decision.
Working context
Temporary whiteboardThe current conversation, active task, loaded skills, retrieved evidence and recent tool output.
Persistent memory
Every sessionA tiny curated set of durable facts about Brad, preferences and operating boundaries.
Skills and references
On demandReusable procedures and focused chapters loaded only when their trigger applies.
Evidence and state
Durable retrievalSession transcripts, SQLite, D1, R2, checkpoints and source snapshots. Queried rather than injected wholesale.
Large organized knowledge
Thousands of Markdown references, database records and source snapshots are manageable when indexed and retrieved selectively. They do not consume the context window merely by existing.
Always loading everything
Huge persistent-memory files, giant main skills and raw project transcripts waste context. The solution is progressive disclosure, checkpoints and ranked retrieval—not deleting useful evidence.
02
Shared vocabulary
These are the exact terms Brad and Sterling should use.
The finite working space sent to the model for the current turn.
System instructions, conversation, loaded knowledge, retrieved evidence and tool results.
MEMORY.md and USER.md: compact, curated and always loaded.
Identity, safety, tools, environment, memory/profile and the skill catalog.
Files such as AGENTS.md or .hermes.md that define a project’s conventions.
An on-demand operating contract loaded when a task matches its trigger.
A separate linked file loaded only when that specific capability is needed.
Durable history in Hermes state.db; searchable, but not loaded wholesale.
A structured, lossy replacement for older turns inside a long session.
Current state, proof, blockers, approvals and the exact next action required to resume.
SQLite, D1, R2 or indexed files holding cases, sources, dates and architecture implications.
Separate configuration, memory, sessions, skills, cron jobs, credentials and gateway state.
03
The working context
What actually enters the model’s whiteboard and what stays outside it.
System prompt+Persistent memory+Conversation+Loaded skills+Retrieved evidence=Current working context
Automatically present
- Sterling’s identity and safety requirements
- Available tool definitions
- Persistent Memory and user profile
- Names and descriptions in the skill catalog
- The active conversation or compacted version
- Relevant project context discovered from the working directory
Not automatically present
- Every local file on the Mac
- Full bodies of every skill
- Every Cloudflare case study
- The complete session database
- Every Gmail, Drive or Google Chat message
- Raw source snapshots and large evidence stores
Important: A file can be enormous on disk without affecting context until it is loaded. Storage capacity and context capacity are separate systems.
04
Persistent Memory
The small curated layer brought into every session.
MEMORY.md
Stable system and operating facts: approved boundaries, established architecture, recurring conventions and durable environment details.
Belongs hereStable preferencesIdentity rulesAccount-routing factsLong-lived safety boundaries
Does not belongTask progressTemporary blockersRaw researchIDs that will soon become stale
USER.md
Stable facts about Brad: response preferences, approval expectations, naming conventions and durable working style.
PurposePrevent repeated correctionPreserve preferred interactionKeep Brad-facing output consistent
ConstraintIt is intentionally small because every character is paid for on every session.
Near-full memory is not lost intelligence. It means completed or procedural material should be moved to skills, checkpoints or evidence libraries so the always-loaded layer remains concise.
05
Skills and deep knowledge
The main skill is a router. Focused references and indexed evidence contain the depth.
Triggers, boundaries, taxonomy and retrieval instructions.
Ranked records are retrieved for the active question. The entire library is never loaded at once.
Main skill
Short operating contract. Target approximately 15–20 KB. A main skill over 50 KB is a strong signal to split it.
Reference file
One decision boundary or capability. Loaded separately only when needed.
Database
Large, changing evidence. Queried by date, keyword, capability, job and architecture relevance.
06
Compaction and continuity
How Hermes protects a long conversation—and why durable checkpoints still matter.
Opening exchangePreserved
Older middle turnsPruned + summarized
Recent tailPreserved
What compaction does
- Prunes large old tool results where safe.
- Protects the initial exchange and recent messages.
- Summarizes goals, constraints, progress, blockers, decisions, files and next steps.
- Updates the prior summary on later compactions.
What compaction cannot guarantee
- Every minor detail survives.
- Project state becomes permanent memory.
- External systems are still in the state described.
- A summary replaces source-of-truth proof.
Before switching initiativesWrite checkpoint → preserve proof → record blocker → identify next command
Compaction is continuity protection. The checkpoint is the durable project ledger.
07
Capacity and scaling decisions
Add resources for measured constraints or durable ownership—not merely because the knowledge library is large.
SignalCorrect responseNot the response
Many Markdown/reference filesIndex and retrieve selectivelyDo not buy another model account
Persistent Memory near its small limitPrune and promote procedures to skillsDo not discard the evidence library
Independent programs need simultaneous ownershipCreate another Hermes profileDo not mix all state into one agent
Sustained provider throttling or quota exhaustionEvaluate another provider account or credential poolDo not infer from one transient timeout
Deterministic high-volume jobsMove execution/state to CloudflareDo not spend model context on repetitive work
Use for isolated ownership
Separate memory, skills, sessions, cron, credentials, safety boundaries or public routes. A dedicated Marketing Machine builder is a logical future profile when continuous ownership competes with Sterling’s executive operations.
Use only for measured provider limits
Trigger on sustained throttling, quota exhaustion, unacceptable concurrency latency, security isolation or contractual requirements—not storage, Markdown count or ordinary compaction.
08
Cloudflare’s role
Cloudflare scales deterministic execution and durable systems. It does not replace the Hermes reasoning process or model context.
Reasoning and governance
- Interpret Brad’s goals
- Research and synthesize
- Choose tools and procedures
- Coordinate specialist agents
- Apply approvals and safety boundaries
- Explain decisions
↔
Deterministic execution
- Workers and Workflows
- Queues and retries
- Durable Objects and locks
- D1 and R2 state/evidence
- Scheduled jobs and APIs
- Sanitized observability dashboards
AI Gateway boundary: AI Gateway observes only calls intentionally routed through it. It does not automatically see the current ChatGPT/Codex session. Local Hermes metrics and AI Gateway metrics are complementary—not interchangeable.
09
Capacity metrics
Sanitized operational measurements from the recurring local audit. No prompts, message bodies, memory text, credentials, internal paths or raw logs are published.
Latest auditLoading…
Overall statusLoading
Audit cadenceMon / Wed / Fri · 11:45 PM ET
Active capacity signals
Loading current audit…
Scaling recommendation
Loading current recommendation…
Publication safeguard
The full audit remains local on the Mac. This page receives only a strict aggregate projection. A manifest lock prevents the recurring metrics publisher from deploying if any unrelated Brad–Sterling2 file changes.