Conversation Logs and Persistent Context

Relevant source files

• Google/Antigravity/Fast Prompt.txt

Purpose and Scope

This document details Antigravity's persistent context system, specifically focusing on Conversation Logs and Artifacts as raw historical data storage. Antigravity maintains two complementary mechanisms for accessing information from past conversations:

1. Conversation Logs and Artifacts (this page): Original, unprocessed conversation data stored in the file system
2. Knowledge Items (KIs) (see 3.2): Distilled, curated knowledge on specific topics

This dual system allows Antigravity to retrieve context at different granularities: reading conversation logs provides complete historical detail for specific conversations, while Knowledge Items provide synthesized knowledge across multiple conversations on related topics.

Sources: Google/Antigravity/Fast Prompt.txt216-221

---

Conversation Logs File System Architecture

Conversation logs are stored in a structured file system hierarchy within the .gemini directory. Each conversation is assigned a unique Conversation ID (UUID format) and maintains its own directory with system-generated logs and artifacts.

Directory Structure

Diagram: Conversation Logs File System Hierarchy

Sources: Google/Antigravity/Fast Prompt.txt204-224 Google/Antigravity/Fast Prompt.txt277-281

File System Components

Component	Path Pattern	Purpose
Brain Root	C:\Users\Lucas\.gemini\antigravity\brain\	Root directory for all conversation data
Conversation Directory	brain/[UUID]/	Container for single conversation's data
System Generated	brain/[UUID]/.system_generated/	System-managed files including logs
Logs Directory	brain/[UUID]/.system_generated/logs/	Contains overview and task-specific logs
Overview Log	logs/overview.txt	High-level summary of conversation
Task Logs	logs/task_NNN.txt	Detailed logs for specific tasks
Artifacts Directory	brain/[UUID]/artifacts/	ASSISTANT-generated content from conversation

Sources: Google/Antigravity/Fast Prompt.txt222-224 Google/Antigravity/Fast Prompt.txt277-281

---

Access Pattern: Staged Reading Approach

Antigravity implements a staged reading approach to efficiently navigate conversation logs without overwhelming context. This pattern prioritizes high-level overview before diving into detailed task logs.

Diagram: Staged Reading Workflow

The staged approach prevents unnecessary context loading:

1. List Directory: Identify available files in .system_generated/logs/
2. Read Overview: Parse overview.txt to understand conversation scope
3. Evaluate Relevance: Determine if detailed logs are needed for current task
4. Conditional Task Logs: Only read task_NNN.txt files if overview indicates relevance
5. Artifacts: Access artifacts directory only if specific files are referenced

Sources: Google/Antigravity/Fast Prompt.txt277-281 Google/Antigravity/Fast Prompt.txt238-239

---

When to Use Conversation Logs

Conversation logs should be accessed when the agent needs specific details from a small number of conversations. This is appropriate when context can be narrowed to particular conversation instances.

Decision Tree for Conversation Log Access

Diagram: Decision Tree for When to Use Conversation Logs

Sources: Google/Antigravity/Fast Prompt.txt225-231

Specific Use Cases

Scenario	Approach	Tools
@mention with Conversation ID	Direct access if information is likely relevant	list_dir, view_file
User mentions "yesterday's conversation"	Identify by recency from available conversation IDs	list_dir, grep_search on overview files
User references specific detail	Search filesystem to identify conversation(s)	codebase_search, grep_search
Need exact wording/code	Access conversation logs for verbatim content	view_file on task logs and artifacts

Sources: Google/Antigravity/Fast Prompt.txt225-231

---

When NOT to Use Conversation Logs

Antigravity explicitly defines anti-patterns where conversation logs should not be accessed to prevent context overload and redundant work.

Anti-Pattern Decision Matrix

Diagram: Anti-Patterns for Conversation Log Access

Sources: Google/Antigravity/Fast Prompt.txt232-239

Explicit Prohibitions

1. When Researching Topics: Always search Knowledge Items first. Only read conversation logs if no relevant KIs exist.

   • Rationale: KIs aggregate knowledge across conversations, providing more comprehensive coverage than individual logs

2. Referenced but Irrelevant: If a KI or another conversation references a conversation ID, but the summary indicates irrelevance to current context, do not access it.

   • Rationale: Prevents following tangential reference chains

3. After Reading Overview: Once overview.txt is read and assessed as not relevant, do not proceed to task logs or artifacts.

   • Rationale: Respects the staged reading pattern and avoids wasting context

4. Broad Knowledge Queries: For questions spanning multiple conversations or general topics, use Knowledge Items exclusively.

   • Rationale: KIs are specifically designed for distilled, cross-conversation knowledge

Sources: Google/Antigravity/Fast Prompt.txt232-239

---

Comparison: Conversation Logs vs Knowledge Items

The persistent context system maintains both raw logs and distilled knowledge. Understanding when to use each system is critical for efficient context gathering.

System Comparison Table

Aspect	Conversation Logs	Knowledge Items (KIs)
Content Type	Raw, original conversation data	Distilled, curated knowledge
Scope	Single conversation instance	Cross-conversation, topical
Update Mechanism	Created during conversation	Generated by KNOWLEDGE SUBAGENT post-conversation
Storage Location	brain/[UUID]/.system_generated/	knowledge/[KI-ID]/
Primary Use	Specific conversation details	Research and knowledge lookup
Access Pattern	Staged reading (overview → tasks)	Check summaries first, then artifacts
Best For	Debugging specific past decisions, retrieving exact code/wording	Understanding architecture, patterns, best practices

Sources: Google/Antigravity/Fast Prompt.txt216-221 Google/Antigravity/Fast Prompt.txt240-249

When to Prefer Each System

Diagram: Decision Criteria for Context System Selection

Sources: Google/Antigravity/Fast Prompt.txt244-249

---

Practical Usage Examples

Example 1: User References Recent Conversation

Scenario: User says "The refactoring we just did to extract the game logic into a separate module broke the unit tests."

Diagram: Example Workflow for Implicit Conversation Reference

Sources: Google/Antigravity/Fast Prompt.txt270-293

Example 2: Multiple KIs Required (Avoiding Conversation Logs)

Scenario: User requests "Add a new AI player to my tic-tac-toe game that uses minimax algorithm and follows the existing game architecture patterns."

Agent receives KI summaries at conversation start:

• game_architecture_patterns KI with artifacts: architecture_overview.md, implementation_patterns.md, class_diagram.md
• randomized_ai_implementation KI with artifacts: random_player.md, ai_player_interface.md, testing_strategies.md
• database_schema KI (irrelevant)
• ui_components KI (irrelevant)

Agent behavior:

Diagram: KI-Only Usage Pattern (No Conversation Logs)

Key observations:

• Agent does NOT call list_dir or grep_search to find conversations
• Agent uses KI summaries already provided in context
• Agent focuses only on relevant KIs, ignoring irrelevant ones
• Conversation logs are never accessed because KIs provide sufficient information

Sources: Google/Antigravity/Fast Prompt.txt255-269

Example 3: No Context Access Needed

Scenario: User asks "What's the difference between async and await in JavaScript?"

Agent behavior:

• Does NOT check conversation logs
• Does NOT search Knowledge Items
• Responds directly with general knowledge
• Rationale: Question does not require project-specific context

Sources: Google/Antigravity/Fast Prompt.txt297-299

---

Tool Integration for Conversation Log Access

File System Navigation Tools

Tool	Purpose	Usage Pattern
list_dir	List contents of conversation logs directory	Called on .system_generated/logs/ to discover available files
view_file	Read specific log or artifact files	Used for overview.txt, task logs, and artifact files
codebase_search	Semantic search across file system	Find conversations by topic when ID unknown
grep_search	Exact pattern matching	Search for specific strings in overview files
find_by_name	Glob-based file discovery	Locate conversations by pattern matching

Sources: Google/Antigravity/Fast Prompt.txt277-281 Google/Antigravity/Fast Prompt.txt338-592

Example Tool Call Sequence

For reading a known conversation (ID: 1a2f082d-72a2-b281-0081-8b9cad0e1f20):

1. list_dir(
     DirectoryPath="C:\\Users\\Lucas\\.gemini\\antigravity\\brain\\1a2f082d-72a2-b281-0081-8b9cad0e1f20\\.system_generated\\logs"
   )
   
2. view_file(
     AbsolutePath="C:\\Users\\Lucas\\.gemini\\antigravity\\brain\\1a2f082d-72a2-b281-0081-8b9cad0e1f20\\.system_generated\\logs\\overview.txt"
   )
   
3. [If relevant] view_file(
     AbsolutePath="C:\\Users\\Lucas\\.gemini\\antigravity\\brain\\1a2f082d-72a2-b281-0081-8b9cad0e1f20\\.system_generated\\logs\\task_001.txt"
   )

Sources: Google/Antigravity/Fast Prompt.txt413-418 Google/Antigravity/Fast Prompt.txt572-581

---

Key Principles Summary

Principle	Description
Staged Reading	Always read overview.txt before task logs; stop if not relevant
KIs First	For topic research, check Knowledge Items before conversation logs
Conversation Logs for Details	Use logs when need specific conversation details or small number of conversations
Respect Irrelevance	If overview indicates irrelevance, do not read task logs or artifacts
Parallel Access	When multiple files needed, use parallel view_file calls where appropriate
Absolute Paths Only	All file operations must use absolute paths (Windows format with backslashes)

Sources: Google/Antigravity/Fast Prompt.txt216-301