Pgramma

中文文档

Pgramma is a local-first digital persona engine: one .pgram file stores identity, memory, and conversation history.

Core Idea

• One file, one persona.
• Memory is append-only and not rewritten.
• Each turn follows: recall -> respond -> evaluate -> store.

Quick Start

Prerequisites

• Rust 1.80+ (edition 2024)
• At least one usable LLM provider endpoint

Setup

git clone https://github.com/KercyDing/pgramma.git
cd pgramma
cp config.example.toml config.toml
# Edit config.toml: set active_provider, model, api_key
cargo run

Commands

Command	Description
/help	Show commands
/test <path>	Seed from file and verify recall
/stats	Show engram and episode counts
/quit	Exit

Configuration

Use config.example.toml as the template and keep local edits in config.toml.

Key points:

• llm.active_provider selects provider.
• llm.providers.<provider>.model and api_key are required for the active provider.
• embedding.model_id controls local embedding model.
• config.toml is ignored by git.

If required fields are missing, startup prints a fatal message and exits gracefully.

Architecture

• src/main.rs: REPL and app bootstrap
• src/app/: bootstrap, REPL loop, and command orchestration
• src/config/: TOML config parsing
• src/domain/: domain models and error types
• src/db/: redb-based storage
• src/llm/: multi-provider inference and evaluator
• src/memory/: embedding and recall pipeline

Testing

cargo test

License

GPL-3.0