Four memory layers.
Zero vendor lock-in.

The schema designed for Layer 1 determines whether Layer 4 is a clean extension or a painful refactor. We designed for Layer 4 from day one.

Memory Architecture

From agent isolation to federated intelligence

L1

Agent Hot Memory

Shipping
  • Per-agent conversation isolation
  • Seven named commanders from persona templates
  • Each with unique soul, tools, skills, permissions, vault access
  • Commander knowledge notes — persistent per-commander memory
  • Schema: conversations.persona column, GetActiveConversation filters
L2

Cold Memory (Global)

Building
  • All conversations from all agents, indexed and searchable
  • Topic-tagged, entity-linked, temporally ordered
  • Cross-agent queries: "what did the strategist decide about X?"
  • Subject to agent VaultAccessGrants
  • Schema ready: topics, entities_mentioned, decisions_referenced columns
  • Memory Manager indexing and retrieval in progress
L3

Team Hot Memory

Shipping
  • Commander Room: multi-agent governance with persistent transcripts
  • Commanders take turns discussing topics from their domain expertise
  • Real-time turn-by-turn delivery via Telegram
  • Schema: conversations.team_id links to teams table
  • Token tracking, duration metrics, structured conversation data
L4

Federated Agent Memory

Foundation built
  • Document attestation chain: SHA-256 hash-linked records
  • Session provenance chain: every agent interaction auditable
  • Proof PDFs with QR codes for offline verification
  • Revocation support — attestations can be formally revoked
  • Full federation (Holochain membranes, DHT discovery) post-funding

"The schema designed for Layer 1 determines whether Layer 4 is a clean extension or a painful refactor."

Schema

How the layers stack

L4
Attestation chain, provenance chain → Holochain membranes, DHT
L3
Commander Room, team_id, shared context, multi-agent consensus
L2
memory_index: source_agent, topic, entity, timestamp
L1
conversations.persona, knowledge notes, soul.md, GetActiveConversation
Each layer extends the schema below it — never breaks it.

Security

Sovereign by architecture, not by promise

We don't ask you to trust us. We built a system where trust isn't required.

Compile-time plugin system

Deny-by-default permissions. No runtime marketplace. ClawHavoc-proof.

PII encryption gate

Automatic PII detection between vault and LLM. 11 scanner patterns, vault-gated redaction.

Privacy classification per channel

CLI = maximum trust. Telegram = standard. Each channel enforces its own redaction level.

Provenance chain

SHA-256 hash chain on agent-audit.db. Every agent interaction is auditable with tamper detection.

Zero telemetry guarantee

No data leaves your machine unless you explicitly route to a cloud model.

VaultAccessGrants

Per-agent, per-vault permissions. Agents only see what they're granted. Persona grants control tool access.

Origin

The 1988 vision

Amiga 500. 2400 bps modem. Autonomous software entities negotiating across phone lines.

In 1988, a kid in Mudgee imagined software agents that could act on your behalf — making deals, sharing intelligence, maintaining boundaries — across the nascent networks connecting home computers. The vision was clear even if the infrastructure wasn't ready.

Thirty-eight years later, the infrastructure caught up. Large language models provide the reasoning. Holochain provides the agent-centric substrate. Go provides the systems-level performance. And the schema designed for Layer 1 was designed with Layer 4 in mind from the very first commit.

Neural Commander isn't a startup idea discovered through market research. It's the culmination of a vision that's been refining itself for nearly four decades — waiting for the technology to catch up to the architecture.

Designed for sovereignty.
Built for permanence.