Echo

Echo is a local persistent memory system for AI agents and developers, built in Go with SQLite, FTS5, DuckDB, and the Model Context Protocol. It gives stateless LLM sessions a durable memory layer that can be queried by an AI host over MCP or managed directly from a terminal.

The project focuses on fast local recall, explicit memory governance, and safe human-in-the-loop maintenance. Memories are stored locally, scoped by context, indexed for keyword search, and exposed through both a JSON-RPC MCP server and a CLI.

Project Portal | Full Documentation

---

Highlights

Area	What it demonstrates
Persistent memory	SQLite stores durable memories with context keys, entry types, tags, source attribution, and reinforcement through UPSERTs.
AI integration	The MCP server exposes store, recall, search, update, deletion, and analytics tools for agent workflows.
Human control	The CLI supports store, recall, search, delete, maintenance, and table, JSON, or CSV output formats.
Search performance	SQLite FTS5 indexes memory content for keyword search, with a short-query fallback for substring matching.
Safety	Destructive operations use an ID-based retrieve, confirm, act flow to avoid accidental deletion by ambiguous content.
Analytics	Telemetry can be written to JSONL and synchronized into DuckDB for context-level usage, cost, hit-rate, and carbon estimates.

---

Architecture

Echo uses a dual-interface architecture. The CLI and MCP transport layer share the same service and persistence code, so AI-driven and human-driven workflows operate on the same local data model.

Path	Use case	Flow
MCP path	AI agent stores or retrieves memory	Agent host -> JSON-RPC over stdio -> MCP tools -> service layer -> SQLite and FTS5
CLI path	Human reviews or curates memory	Terminal command -> CLI dispatcher -> service layer -> SQLite and FTS5
Analytics path	Usage analysis and memory tuning	Telemetry JSONL -> DuckDB sync -> context-level analytics

graph TD
    subgraph "Interfaces"
        A[AI Agent or IDE] -- "JSON-RPC over stdio" --> B[MCP Transport]
        H[Terminal] -- "CLI Commands" --> I[CLI Dispatcher]
    end

    subgraph "Core Logic"
        B --> C[Shared Service Layer]
        I --> C
        C --> D{Storage Engine}
    end

    subgraph "Persistence and Analytics"
        D --> E[(SQLite Memories)]
        D --> F[(FTS5 Search Index)]
        D --> G[(DuckDB Analytics)]
    end

Loading

---

Tech Stack

Layer	Tools
Language	Go
Data stores	SQLite, FTS5, DuckDB, JSONL telemetry log
Observability	Memory telemetry events, DuckDB analytics views, benchmark suite
Testing	Go testing package, table-driven tests
CI/CD	GitHub Actions

---

Documentation

• Full Documentation
• Architecture
• Decision Records

---

Demos

• MCP Server: AI Agent Interface
• CLI Interface: Human Usage

---

Local Setup

Build and verify the project:

make build
make test
make bench

Install the CLI:

make install

Use the CLI directly:

echo-cli store -content "Use FTS5 for memory search." -context project:echo -type directive -tags search,sqlite
echo-cli recall -contexts global,project:echo -limit 10
echo-cli search -query FTS5
echo-cli maintain -rebuild

Run the binary as an MCP server by launching it from an MCP host with stdin connected to the JSON-RPC transport. When no CLI subcommand is provided and stdin is not a terminal, Echo starts the MCP server automatically.