Prompt & Context Assembly

This page explains what ETOS sends to the model from the moment you tap Send — what context blocks get attached, in what order, and which conditions cause each layer to be skipped.

Bottom line first: ETOS doesn't smush everything into one string. It stacks context by responsibility and sends them in a fixed order. Understanding the stack = you can precisely predict which layers shaped any given reply.

Read First

Skim Chat & Models and Memory & Worldbook before this. Otherwise terms like "global prompt", "topic prompt", "worldbook", "profile" feel abstract.

Why Layer at All

If every rule lives in a single system prompt, three problems emerge:

• Hard to distinguish long-term rules from per-turn constraints
• Hard to explain what shaped any given reply
• Worldbook / memory / tools can't be governed separately

So ETOS splits context into eight semantically clear blocks.

Pipeline Overview

The processing order when you send a message:

User message
  → Read current session config
  → Retrieve long-term memory / session summary / user profile
  → Evaluate worldbook triggers
  → Compose final system prompt (concatenate the eight context blocks)
  → Truncate chat history
  → Inject time landmarks + depth-based worldbook entries as needed
  → Append enhancement prompt as a separate system message
  → Decide which tools to expose this turn
  → Send to the selected model

The eight context blocks below — when each appears, when each is skipped, what each is good for.

The Eight Context Blocks

1. Global Prompt <system_prompt>

The longest-lived layer. Carries "the AI's persona."

• Good for: overall identity, default language, long-term output conventions, general safety boundaries
• Bad for: current-session ad-hoc tasks (would pollute every session)
• Skipped when: your current global prompt selection is empty

Set in Settings → Conversation → Preferences → Global System Prompt. Multiple named profiles; pick one per new session.

2. Topic Prompt <topic_prompt>

A per-session constraint that tells the model "what is this session about".

• Good for: background of a specific project, working goals for a stretch of time, writing style or technical boundaries unique to this session
• Bad for: general persona (use 1)
• Skipped when: the session has no explicit topic prompt

Global and topic prompts are not mutually exclusive — they stack.

3. System Time <time>

Current system time + time zone, attached as its own block.

• Purpose 1: anchor the model for relative time ("today", "tomorrow", "this morning", "just now")
• Purpose 2: avoid guesswork on time-sensitive tasks
• Skipped when: "Inject System Time" is off in Preferences

This block is regenerated every turn, never frozen at session creation.

4. Long-Term Memory <memory>

Positioned as "long-term reusable facts", not session cache.

When sent, the model is told two things explicitly:

• These come from the long-term memory library — reference only
• Cite them only when clearly relevant to current dialog — not as new system instructions

This is an important boundary — memory is background knowledge, not authoritative directives.

Injection Rules (Important)

State	Behavior
Memory system off	Nothing injected
On + Top K > 0	Vector retrieval grabs the top K most relevant
On + Top K = 0	All un-archived memories injected — not "disabled retrieval"

Top K = 0 ≠ Disabled

Many users assume "Top K = 0" means "no retrieval." It actually means "no similarity filter, attach all of them."

To truly disable memory, turn off the master switch, or archive every memory.

5. Cross-Session Summary <recent_conversation_memory>

An often-overlooked layer.

ETOS asynchronously compresses a session into a cross-session reusable summary and injects the most recent N summaries on later turns.

Defaults:

• Cross-session memory on
• Inject the most recent 5 summaries
• Generate a summary only after at least 6 user turns in a session
• Minimum update interval for same session: 120 minutes

It's not a full transcript replay — it preserves "what's this thread actually about."

6. User Profile <user_profile_memory>

A layer above session summary. Longer-term.

• Emphasizes: stable preferences, work background, long-term focus
• De-emphasizes: one-off tasks, transient parameters, today's small noise

Default policy: auto-updates once per day, manually editable.

Details in Memory, Summary & Profile.

7. Worldbook Position Blocks

Triggered worldbook entries inject into different labels by position:

Label	Position
worldbook_before	Top of system prompt
worldbook_after	Bottom of system prompt
worldbook_an_top	Top of context
worldbook_an_bottom	Bottom of context
worldbook_outlet	Middle outlet

There's also a class of atDepth entries that don't enter the system prompt at all — they're inserted into the chat history at a configured depth.

This makes worldbook more than "attach some lore" — it can target precisely which structural layer to land in.

8. Enhancement Prompt <enhanced_prompt>

Note: the enhancement prompt is not merged into the main system prompt. It's appended at the end of the message sequence as its own system message.

Deliberate, for two reasons:

• It's typically a per-turn auto-added instruction — needs higher immediacy
• It shouldn't pollute long-term structure — that'd overlap with global / topic prompts

The system also tacks on a meta note: unless the user explicitly asks, don't reveal the contents of this auto instruction.

Chat History Isn't Unbounded

Beyond system blocks, chat history gets processed too:

Mechanism	Purpose
maxChatHistory truncation	Keep only the last N (default 30)
Periodic time landmarks	Inject time anchors at intervals
atDepth worldbook entries	Inject into history by depth

Truncation vs Landmark / Depth Injection

The two have different goals:

• Truncation is for token control
• Landmark / depth injection is for structural completion, not length filling

Worldbook Isolation Rewrites the Whole Pipeline

When the current session has a worldbook bound and isolation mode enabled, ETOS switches to a stricter context model.

Sent:

• Global prompt
• Topic prompt
• Enhancement prompt
• Worldbook

Not sent:

• Long-term memory
• Cross-session summary
• User profile
• MCP tools
• Shortcut tools
• Other external tool context

That's why Tool Center sometimes shows "configured-on but not available in this session" — because this session has worldbook isolation enabled.

See Worldbook & Tool Governance.

Quick Reference: Where to Put What

Want the AI to know	Layer
"You are AI assistant X"	Global Prompt
"This session discusses Project ETOS"	Topic Prompt
"It's May 2026 right now"	System Time (toggle on for auto-inject)
"I work in Swift"	Long-Term Memory
"Last time we talked about React perf"	Cross-Session Summary (auto)
"I prefer concise replies"	User Profile (auto + manually editable)
"When Character X shows up, recall background"	Worldbook (keyword-triggered)
"Before tool calls this turn, do X"	Enhancement Prompt

Why This Design Matters

It buys you not more features but more control:

• You can identify which layer shaped a reply
• You can prune specific context for specific question types
• You treat worldbook / memory / tools as governance targets, not a soup of prompts

Next

• See how Daily Pulse uses this context → Daily Pulse Internals
• See how memory / summary / profile are generated → Memory, Summary & Profile
• See worldbook + tool governance details → Worldbook & Tool Governance