AgentGo - Local RAG System with Cognitive Memory

AgentGo is a local-first RAG + Agent framework that features an evolving memory layer. It transitions from raw data points to high-level insights via autonomous reflection.

When to Use

• Build local knowledge base from documents
• AI assistants with evolving long-term memory (Hindsight)
• Reasoning-based retrieval for long documents (PageIndex)
• Q&A systems backed by your own data
• AI agents with MCP tools and Skills
• Multi-provider LLM (OpenAI, Anthropic, Ollama)

Quick Start

As a Library

package main

import (
    "context"
    "fmt"
    "github.com/liliang-cn/agent-go/pkg/agent"
)

func main() {
    ctx := context.Background()

    svc, _ := agent.New("my-agent").
        WithPrompt("You are a helpful assistant.").
        WithRAG().
        WithMemory(). // Enables Cognitive Memory (Facts -> Observations)
        WithDebug(false).
        Build()
    defer svc.Close()

    // Conversational chat with memory recall
    result, _ := svc.Chat(ctx, "I am a Go developer.")
    fmt.Println(result.Text())

    // Memory visibility: check what the agent "knows" and why
    for _, m := range result.Memories {
        fmt.Printf("[%s] %s (Confidence: %.2f)\n", m.Type, m.Content, m.Confidence)
    }
}

CLI Usage

# Ingest documents
agentgo ingest ./documents/

# Chat with memory transparency
agentgo chat "What do you know about my projects?" --show-memory

Core Features

1. Cognitive Memory (Hindsight)

AgentGo doesn't just store facts; it learns patterns.

• Facts: Atomic data from chat.
• Observations: Synthesized insights from multiple facts via Reflect().
• Traceability: Track the evolution from raw fact to high-level belief.

2. Reasoning Retrieval (PageIndex)

For massive documents or when embeddings are weak:

• Index Navigator: LLM-driven tree search over hierarchical summaries.
• Hybrid Search: Parallel Vector + Index reasoning fused via RRF.

3. RAG Operations

Ingest docs → chunk → embed → SQLite vector store → hybrid search.

4. Multi-Provider LLM

OpenAI, Anthropic, Ollama, DeepSeek, and Azure.

5. Skills & MCP

Extensible via YAML+Markdown Skills or standard MCP tool servers.

Architecture

┌──────────────────────────────────────────────┐
│                AgentGo System                   │
├──────────────────────────────────────────────┤
│  RAG Store │ LLM Pool │ Cognitive Memory     │
│       ┌────┴──────────┴────────┐             │
│       │      Agent Service     │ ← Reflect   │
│       │  ToolRegistry (Unified)│   Engine    │
│       │  Index Navigator       │ ← PageIndex │
│       └────────────────────────┘             │
└──────────────────────────────────────────────┘

Builder Pattern

svc, _ := agent.New("name").
    WithRAG().
    WithMemory().
    WithMemoryReflect(5).      // Auto-reflect after 5 new facts
    WithMemoryHybrid().       // Enable Vector + Index search
    WithMemoryBank("mission", []string{"directive1"}).
    Build()

Invocation API

Method	Session	Cognitive Memory
Ask(ctx, p)	No	Implicit recall
Chat(ctx, p)	Yes	Fact extraction + Recall
Run(ctx, g)	Optional	Full evolution + Reflect

See Also

• API Reference - Detailed API
• Architecture - Design decisions
• Configuration - Config options