Skip to content

Latest commit

 

History

History
59 lines (45 loc) · 1.21 KB

memory.md

File metadata and controls

59 lines (45 loc) · 1.21 KB

Memory System

Overview

AlphaLLM features a dual memory system combining short-term and long-term memory for contextual conversations.

Short-Term Memory (STM)

Storage

  • Uses TinyDB for local JSON storage
  • Stores conversation messages per user/channel
  • Automatic cleanup (max 50 messages per conversation)

Structure

{
  "user_id": 123456789,
  "conv_id": "channel_id",
  "content": "Hello world",
  "role": "user",
  "timestamp": 1706745600.0
}

Retrieval

  • Fetches last N messages for context
  • Configurable via MAX_CONVERSATION_HISTORY

Long-Term Memory (LTM)

Storage

  • Uses Faiss for vector similarity search
  • Stores summarized memories and documents

Retrieval

  • Semantic search based on input similarity
  • Reranked using LightReranker
  • Configurable via MAX_LTM_RESULTS

Memory Integration

In Prompts

  • STM provides recent conversation context
  • LTM provides relevant historical information
  • Combined for rich, contextual responses

Automatic Management

  • Messages stored after each response
  • Periodic summarization (planned)
  • Memory limits prevent overflow

Configuration

[limits]
max_stm_messages = 50
max_ltm_results = 5
max_conversation_history = 10