Inference layerShipped · validated
AI models that remember.
Stateful inference infrastructure for Mamba, Mamba2, and hybrid SSM models. Save model state after a turn, restore in two milliseconds, process only new tokens.
Explore Engram
Memory at every layer.
AI forgets at every layer of the stack. Engram restores model memory between inference turns. Plexium gives coding agents durable knowledge of your codebase. Markedup turns markdown files into a knowledge graph. Three products, one thesis: AI should remember.
Knowledge layerIn development
Plexium.
Agents that know your codebase.
Repository memory for coding agents. A persistent, agent-maintained wiki that compounds with every commit, replacing the rediscover-from-scratch pattern that burns tokens and context.
Explore PlexiumFilesystem layerIn development
Markedup.
Your markdown is a knowledge graph.
Filesystem-native knowledge graph built from markdown frontmatter. No external database. Your notes become a queryable memory layer for agents and humans alike.
Explore MarkedupNext in the stack
Memory at the edge.
Synapse is our llama.cpp fork for ARM NPU hardware. Today it focuses on performance optimization — running frontier models on devices cloud GPUs can't reach. Snapshot-style persistent memory is on the roadmap, bringing Engram's stateful inference to the edge.
Explore Synapse
Edge inference runtime
llama.cpp fork · ARM NPU