In-Depth Look at the Engram Memory System

The 'cold start' problem, where previous decisions or context must be re-explained every time an LLM session begins, is a major factor that undermines user experience. The existing MEMORY.md approach is useful for storing fixed facts, but it has limitations in capacity and structure when it comes to capturing the background of decisions or complex reasoning processes.

Built on Weaviate's vector search technology, Engram is designed to fill this gap in context. Going beyond simple text storage, Engram structures memory by topic so that LLMs can retrieve exactly the information they need at the right time.

Engram organizes memory into the following four main categories:

• communication-style: output format, tone, preferences, etc.
• domain-context: knowledge related to role, company, and product
• tool-preferences: languages, frameworks, and tool stacks
• workflow: the user's way of working

For efficient session management, Engram uses a pattern of loading project context at the start of a session and saving data at key decision points or upon task completion. In particular, to maximize retrieval efficiency, it improves data precision by storing short, focused entries of 2 to 4 sentences per topic.