AgentForge

AgentForge uses a Config singleton that loads and parses YAML files from a .agentforge directory, enabling agents and workflows to be defined declaratively without code changes. The ConfigManager builds structured configuration objects that support dynamic model selection and prompt updates at runtime without restarting the application, using a file-watching pattern for hot-reload capability.

AgentForge provides a Cog class that orchestrates multiple Agent instances in a defined workflow sequence, managing execution order, data flow between agents, and memory context propagation. Cogs are configured via YAML flow definitions that specify which agents run, in what order, and how outputs from one agent feed into the next, with the MemoryManager automatically injecting contextual information before each agent executes.

AgentForge uses Chroma as the default storage backend for all memory types, providing vector-based semantic search capabilities. The integration handles embedding generation, vector storage, and retrieval, enabling agents to find relevant memories based on semantic similarity rather than exact keyword matching. Chroma can be deployed locally or remotely, supporting both development and production scenarios.

AgentForge includes a parsing processor that extracts structured data from agent outputs, handling JSON parsing, regex extraction, and custom parsing logic. The processor enables agents to generate structured outputs (JSON, YAML, etc.) that are automatically parsed into Python objects, with error handling for malformed outputs and fallback strategies.

AgentForge implements a base API layer that abstracts away provider-specific details (OpenAI, Anthropic, Ollama, etc.), allowing agents to be written once and run against any supported LLM without code changes. The framework handles provider-specific API differences, authentication, and model parameter mapping through a unified interface, with model selection configurable per-agent via YAML.

AgentForge implements a MemoryManager that coordinates three distinct memory types: Persona Memory (agent identity/instructions), Chat History Memory (conversation context), and ScratchPad Memory (working state). Each memory type is backed by a pluggable storage backend (Chroma vector DB by default) and is automatically injected into agent prompts before execution, enabling agents to maintain context across multiple invocations without explicit state management.

AgentForge provides an Actions system (note: marked as deprecated in docs but still present) that enables agents to call external functions and tools through a schema-based registry. Tools are defined declaratively with input/output schemas, and the framework handles marshaling arguments from LLM outputs into function calls, with support for multiple tool providers and custom tool implementations.

AgentForge includes a prompt processor that handles template variable interpolation, memory context injection, and prompt formatting. Prompts are stored as templates in YAML files with placeholders for variables, memory content, and dynamic values that are resolved at agent execution time, enabling reusable prompt templates that adapt to different contexts.

AgentForge includes a testing framework that enables developers to write tests for agents, validating outputs against expected results, checking memory state changes, and verifying workflow execution. Tests are integrated with the configuration system so agents can be tested in isolation or as part of larger workflows, with support for mocking external dependencies.

AgentForge implements a logging system that captures agent execution traces, including inputs, outputs, memory state changes, and timing information. Logs are structured and queryable, enabling debugging of agent behavior and performance analysis. The logging system integrates with the configuration system to enable per-agent log levels and output destinations.

AgentForge includes utilities for deploying agents to Discord as bots, handling message parsing, response formatting, and conversation context management. The Discord integration maps Discord messages to agent inputs and agent outputs to Discord messages, enabling agents to be accessed through Discord without additional wrapper code.

AgentForge implements Persona Memory that stores agent identity, behavioral instructions, and role-specific information. Personas are defined in YAML and automatically injected into agent prompts, enabling agents to adopt different roles, communication styles, and expertise areas without code changes. Multiple personas can be swapped at runtime for the same agent logic.