System Prompt

System Prompt Construction

Overview

The system prompt is split into two layers for prompt caching efficiency:

1. Static system prompt — Base rules + identity files (SOUL.md, IDENTITY.md, USER.md). Cached and reused across requests.
2. Dynamic context preamble — Prepended to each user message. Contains date/time, memories, skills, and reply context.

This separation means the LLM can cache the system prompt tokens and only process the changing preamble as new input.

Key Files

File	Role
src/system-prompt.ts	getSystemPrompt(), buildContextPreamble(), invalidateSystemPromptCache()

Static System Prompt

getSystemPrompt(identity?) builds the full system prompt by concatenating:

BASE_SYSTEM_PROMPT

The hardcoded base prompt that defines:

• Role: “You are a personal companion”
• Rules:
  • Be concise (Telegram context)
  • Proactively store memories
  • Search broadly when recalling
  • Confirm time/message before creating reminders
  • Explain what/why before self-editing
  • Never deploy without passing tests
  • Never edit files outside src/, cli/, or extensions/
• Telegram interactions: How to use telegram tools, message ID format
• Proactive communication: When to add context beyond the bare minimum
• Identity files: How to use identity_read/identity_update tools
• Extensions: How tools and skills work, when to call extension_reload

Identity Sections

If identity files are loaded, they are appended as sections:

## Identity
<content of IDENTITY.md>

## User
<content of USER.md>

## Soul
<content of SOUL.md>

Note the order: Identity, User, Soul. The Soul section (personality) comes last so it has the strongest influence on the model’s behavior.

Caching

The system prompt is cached in module-level variables. The cache key is a pipe-delimited concatenation of all three identity file contents. invalidateSystemPromptCache() clears the cache (called by identity_update and extension_reload).

Dynamic Context Preamble

buildContextPreamble(context) creates a text block prepended to each user message. It contains:

1. Context Header

[Context: Monday, February 24, 2026 at 3:15 PM (America/New_York) | telegram]

Includes the current date/time formatted using Intl.DateTimeFormat in the configured timezone, the message source, and a DEV MODE flag if applicable.

2. Dev Mode Warning

[Running in development -- hot reload is active, self_deploy is disabled]

Only present when NODE_ENV=development.

3. Recent Memories

[Recent memories -- use these for context, pattern recognition, and continuity]
- (preference) User prefers dark mode
- (fact) User works at Acme Corp

The 10 most recent memories, regardless of relevance. Gives the agent temporal continuity.

4. Relevant Memories

[Potentially relevant memories]
- (note) Meeting with Bob about the API redesign (87% match)

Up to 5 semantically relevant memories with match percentage. Deduped against recent memories.

5. Active Skills

[Active skills -- follow these instructions when relevant]

### daily-standup
When the user asks for a standup...

Up to 3 skills selected by embedding similarity (threshold 0.35).

6. Reply Context

[Replying to: "what was that thing you mentioned yesterday?"]

When the user replies to a specific message in Telegram, the original message text (truncated to 300 chars) is included.

Date/Time Formatting

Uses Date.toLocaleString() with the configured timezone:

[Current time: 3:15 PM | Monday March 24 2026 | America/New_York | telegram]

Full Prompt Assembly

The complete prompt the LLM sees:

[System prompt]
  BASE_SYSTEM_PROMPT
  ## Identity (if loaded)
  ## User (if loaded)
  ## Soul (if loaded)

[User message]
  [Context: Monday, February 24, 2026 at 3:15 PM (America/New_York) | telegram]
  [Recent memories...]
  [Relevant memories...]
  [Active skills...]
  [Reply context...]

  <actual user message>

The preamble is prepended directly to the user’s message text (no separator).