01 · The Agent
Neo
Neo is the default agent of the Matrix platform and the one you talk to. It runs a staged turn loop: prepare the window, generate, deliberate on what the model actually said, act through tools, then close the turn and write the outcome back to memory.
Every piece of run state is born inside a turn struct, so nothing leaks between conversations unless Neo chooses to remember it. When a step budget runs out, Neo reports an honest partial and never a fabricated success.
- Staged loop: prepare, generate, deliberate, act, close
- Adaptive step budgets scaled by turn complexity
- Evidence-delta stall detection forces a revised plan when work stalls
- Silent-voice controller that can edit its own prior message in place
02 · The Memory
Cortex
Cortex is a persistent memory engine with one Pebble brain per user, append-only by design. Nine typed memory kinds organize experience; HNSW vectors make recall associative; a temporal ladder rolls moments up into days, weeks, and seasons of a working life.
Every journal entry feeds a Merkle mountain range root, so an agent's history is not just stored but provable. A vault of envelope encryption uses platform, user, and per-object keys to keep each brain sealed at rest.
- Append-only journal with MMR and sparse Merkle tree integrity roots
- Nine memory types, from episodic moments to self-model records
- Temporal ladder with automatic rollups from hours to long horizons
- Envelope encryption: KEK → user key → per-object DEKs (AES-256-GCM)
03 · The Executor
MCL
The Matrix Command Language is the second rail. When work becomes consequential, such as moving money, signing a transaction, or changing irreversible state, intent leaves the conversational rail and enters a compiler.
Natural language becomes a typed intermediate representation, then a plan, then a signed walk through that plan, attested at every step. The result is deterministic replay: what the agent meant is what the machine did, and anyone can check.
- Compile → plan → walk → attest, signed end to end
- Typed Intent IR treats intent as data, not prose
- Deterministic replay of every consequential action
- MCP tool subprocesses isolated from the conversational loop