What is neural_mesh?

NEURAL_MESH is a memory system built for AI agents, designed to replace the common approach of dumping everything an agent has learned into one flat, ever growing text file. The README argues that flat memory files cause real problems in production: they become too large for the context window, they mix different kinds of information together so that a one-time event and a stable fact get treated the same way, and outdated facts stick around alongside their updated versions, which can cause an agent to act on stale information. NEURAL_MESH addresses this by separating memory into five distinct types, including facts, one-time events, step-by-step procedures, sensory details, and future intentions, so a search for one kind does not get diluted by another. Instead of storing memories as a simple list, each piece of memory automatically links to related pieces, forming a graph rather than a flat file. When the agent asks a question, the system does not just return the single closest match, it spreads activation across linked memories and returns a related cluster, which the README calls resonance retrieval. Memories are also split into a short-term hot lane and a longer-term cold lane, with a sleep cycle that periodically strengthens useful memories and prunes weak or outdated ones. When a fact changes, the new version marks the old one as superseded, so future searches only surface the current, correct information instead of both old and new answers at once. For large tool outputs, the system stores the content externally and gives the agent a small pointer instead of the full text, keeping the context window from filling up with things like long log dumps. Installing it requires only cloning the repository and running a demo script using Python's standard library, with an optional add-on for real embeddings. The README includes benchmark comparisons against plain vector similarity search, the kind used by tools like Mem0, on the same underlying embeddings. The headline result claims that when facts get updated, plain vector search returned the outdated version most of the time, while NEURAL_MESH consistently returned the current, correct fact. On simpler recall tasks without conflicting information, both approaches performed similarly, which the README reports honestly as a tie rather than overstating the difference.