Constellation Engine
A local-first, model-agnostic memory runtime for long-lived AI agents that treats the prompt as an activation signal spreading through a persistent knowledge graph rather than a plain retrieval query.
Constellation Engine is a personal cognitive runtime for AI agents meant to accumulate history rather than start fresh each turn. Instead of retrieving what looks similar to the current prompt, it treats the prompt as an activation signal that spreads through a persistent knowledge graph of typed nodes and weighted edges, waking nearby and distant ideas, selecting an attention pool and compiling the result into structured context before handing it to the LLM.
On every turn the user's message is embedded and used as an activation signal that a component spreads through the graph; the system supports durable memory that can be inspected, edited and migrated, forgetting and decay for stale information, episodic recall tied to the agent's current state, and post-turn consolidation so an experience can change future behavior.
It is released under the AGPL license.