fleet-agent

A context-aware development assistant for AgenticFleet that maintains persistent memory across sessions using a hybrid NeonDB + ChromaDB architecture.

Memory Architecture

Dual Storage

• ChromaDB (Semantic): Skills, patterns, code snippets with embedding-based search
• NeonDB (Structured): Sessions, users, analytics, skill metadata with SQL queries

Context Layers

1. Core Memory (.fleet/context/core/): Always loaded

   • project.md: Architecture, conventions, tech stack
   • human.md: User preferences, communication style
   • persona.md: Agent guidelines, tone

2. Topic Blocks (.fleet/context/blocks/): Loaded on demand

   • project/: commands, conventions, gotchas, architecture
   • workflows/: git, review
   • decisions/: ADRs

3. Skills (ChromaDB + NeonDB): Semantic + structured patterns

Usage Examples

Learn a Pattern

/fleet-agent learn --name "add_dspy_agent" --category "agent" --content "Create agent via AgentFactory with DSPyEnhancedAgent wrapper..."

Recall Information

/fleet-agent recall "DSPy typed signatures"
/fleet-agent context "add a new agent for web search"

Analyze Code

/fleet-agent analyze src/agents/coordinator.py

Session Management

/fleet-agent session start
/fleet-agent session status
/fleet-agent session summary "Completed agent creation workflow"

Commands

| Command | Description | | ------------------------------------------------------- | ------------------------------ | | learn --name <name> --category <cat> --content <code> | Save pattern to both databases | | recall <query> | Search NeonDB + ChromaDB | | context <task> | Load relevant context blocks | | analyze <file> | Analyze code structure | | session start | Start new session | | session status | Show current session | | session summary <text> | Save session summary | | stats | Show development metrics |

Auto-Learning

Automatically extracts and saves patterns after successful task completion with detailed code examples:

name: pattern_add_dspy_signature
category: dspy
description: How to create a DSPy signature with TypedPredictor
implementation: |
  class TaskAnalysisOutput(BaseModel):
      complexity: Literal["low", "medium", "high"]

  class TaskAnalysis(dspy.Signature):
      task: str = dspy.InputField(desc="Task to analyze")
      analysis: TaskAnalysisOutput = dspy.OutputField()

Implementation

Main script: .fleet/context/scripts/fleet_agent.py

Invocation: uv run python .fleet/context/scripts/fleet_agent.py <command>

Dependencies: neon_memory.py, chroma_driver.py, memory_loader.py

See Also

• memory-system-guide.md: Complete memory system documentation
• .fleet/context/MEMORY.md: Memory hierarchy and commands