Mazemaker

What Mazemaker does

Mazemaker is a local memory engine that treats memory as formation and consolidation rather than retrieval. It turns raw conversations into a graph of atomic facts, semantic links, supersession chains, synthesized abstractions and temporal trajectories, which an agent can walk instead of searching. The repository description names spreading activation, embedding-based recall, autonomous dream consolidation and C++ LSTM plus kNN pattern learning.

The README compares its pieces to an operating system kernel: agents are processes, consolidation and supersession are memory management, the knowledge graph is the filesystem, dream cycles are the scheduler, and federation is inter-process communication. When a newer fact contradicts an older one, the older one is superseded instead of duplicated. It works with MCP agents and persists across sessions. The README reports R@5 of 0.8426 and R@10 of 0.9000 on a 500-question LongMemEval-oracle run with a 25k-memory haystack, along with negative-control tests where removing a mechanism is expected to lower the score.

Key features

• Knowledge graph of atomic facts, semantic links and temporal trajectories
• Dream cycles that consolidate memory in the background while the agent is idle
• Conflict supersession that replaces stale facts instead of duplicating them
• Multi-hop recall across A to B to C reasoning chains
• Embedding-based recall with explanations of why a memory was recalled
• Federation between agents and C++ LSTM plus kNN pattern learning

When to use it

• Keeping an agent's knowledge of your evolving preferences across sessions
• Linking related facts from separate conversations
• Replacing outdated information when your plans or decisions change

Who it is for: Developers who want a persistent, local memory layer for MCP-capable agents, including Hermes Agent.

How it fits with Hermes Agent

The repository description says it serves any MCP client and Hermes Agent, so it is a general memory system with Hermes support rather than a Hermes-only project.

Note: Benchmark figures in the README are self-reported, and GitHub reports the license as NOASSERTION.

FAQ