Knowledge & Memory
Praxis' brain was designed to evolve as needed. Knowledge is stored across two organs: the Cortex, a knowledge graph it consults before every decision, and the shared-mind vault, one durable memory shared across all agents. A Knowledge Campaign learns about the latest topics relevant to Praxis and its projects overnight so that knowledge can be applied in the morning.
The Cortex Knowledge Graph
The Cortex holds a 152,000-node Neo4j graph under a disciplined 9-type ontology with 16 canonical relationship types. Below: the real topic constellation from the console with the graph's largest communities and the bridges between them.
▸ hover a star for its top entities or select a corner station for how the brain works
◉ live snapshot 2026-08-12 · 585 topic communities · 26 largest shown · sized by knowledge count
{{ gDesc }}
The Knowledge Campaign
The machine's failure path, in detail. A blocked task declares a need, the need becomes a gap node in the graph, and the system hunts it down overnight so answers return to execution by morning.
{{ kDesc }}
The Shared-Mind Vault
The Recent Activity ticker on the bridge is this vault changing — every entry is a write to the one memory shared by all agents.
Durable memory lives as git-backed markdown that stores memories, project files, incident post-mortems, and a curated skills library with usage telemetry. Claude Code, Codex, Antigravity, and the local model all read and write the same vault, so a lesson learned in one session is available to every mind in the next.
A watcher daemon re-indexes on every change: the console entries with vault-watcher picking up "1 memory, 2 generateds" are that pipeline running, each write landing in the Cortex's semantic memory within seconds of hitting the disk.
Because it's files in git, memory is human-readable, diffable, versioned, and recoverable. No opaque database of the machine's mind.