beilusaiying/always-accompany

What it solves

always-accompany is designed to solve the problem of "context collapse" and memory loss in long-term AI interactions. It prevents the AI from forgetting important details in large projects or long conversations by implementing a structured memory system that avoids overloading the token window with irrelevant data while ensuring the most pertinent information is retrieved and injected into the prompt.

How it works

The project uses a multi-layered approach to manage context and attention:

• Layered Memory: It organizes data into hot (recent/active), warm (phase-level), and cold (deep history) layers, along with structured Data tables for facts. These are stored as plain JSON or Markdown files for user transparency.
• P1 Memory Recall: A front-loaded retrieval mechanism (P1) analyzes the user's input to find relevant excerpts from the long-term memory layers before the main AI generates a response. This acts as a dynamic attention mechanism outside the model.
• Context Management: It provides granular control over the context window, allowing for file-read-level cleanup and tiered compression to remove stale or redundant information without deleting the permanent record.
• Plugin-Based Framework: Core features are delivered as plugins, allowing the system to be extended with custom JS, Python, or standalone programs.
• Mode-Specific Isolation: Different modes (Smart, Chat, Code, Work) maintain their own memory tables and retrieval routes to prevent cross-project leakage.

Who it’s for

• Power users who want a highly configurable AI companion for roleplay, coding, and general productivity.
• Developers looking for an agentic workbench that can read/modify local files and execute commands with fine-grained permission controls.
• AI enthusiasts who prefer "white-box" transparency where they can inspect exactly what the AI remembers and why it recalled specific information.

Highlights

• Multi-purpose modes: Dedicated environments for Chat/Roleplay (SillyTavern compatible), IDE-level coding, and work productivity.
• Agentic capabilities: Integrated toolchain supporting MCP, file system access, and command execution.
• Desktop integration: Live2D companion with screen awareness and proactive commenting.
• Extensive bot support: Built-in shells for nine different platforms including Discord, Telegram, and Slack.
• Transparent memory: A white-box diagnostic panel to inspect the P1 recall chain and memory retrieval process.