HashMo InteractiveOpen protocols · independent intelligence
AIS ROADMAP v1.0 · FROZEN

An open language for independent intelligence

Different AI systems.One shared protocol.

The AI Interoperability Standard defines how independently built AI systems exchange work, context, permissions, provenance, and outcomes without surrendering autonomy to one vendor, one model, or one central platform.

Explore the eight primitives See how OSC and AIS fit together
Provider neutralUser-controlled permissionsNo chain-of-thought required
OSC plus AIS symbol representing the operating environment and AI interoperability standard
OSCOperating environment+AISOpen interoperability standard

Two layers, one goal

AIS lets systems speak. OSC helps organizations work.

AIS does not prescribe how an intelligence thinks. OSC does not own the protocol. Together they provide an open language and a human-directed place to use it.

01
AIS

AI Interoperability Standard

The shared protocol: identity, envelopes, primitives, permissions, provenance, memory exchange, governance extensions, and conformance.

  • Open and implementation-agnostic
  • Works across providers and deployments
  • Small stable core, optional extensions
02
OSC

Open Source Connector

The operating environment: teams, missions, channels, memory, approvals, artifacts, federation, and the organizational surfaces where interoperable work becomes visible.

  • Human-controlled coordination
  • Persistent organizational context
  • Reference implementation evidence

Standard-0003

Eight primitives. Everything else is composition.

AIS keeps the universal vocabulary intentionally compact. Rich workflows emerge from payloads and sequences instead of an ever-growing list of protocol verbs.

UNIVERSAL MESSAGE ENVELOPE
{
  "spec_version": "1.0.0",
  "message_id": "uuid",
  "primitive": "request",
  "correlation_id": "uuid",
  "from": { "participant_id": "org-a" },
  "to": { "participant_id": "org-b" },
  "payload": { ... },
  "signature": null
}
Exactly one primitive · one sender · one unique message ID
pingReachability
announcePresence & capability
requestAsk for work
responseReturn an outcome
proposalOffer a course
voteRecord a position
artifactPublish durable work
eventAppend telemetry

Protocol invariants

Open at the edge. Strict at the core.

01

Independent by design

Commercial models, open-source systems, enterprise deployments, local assistants, and future architectures remain sovereign.

02

User ownership

Only explicitly authorized context, memory, permissions, and work cross system boundaries.

03

Traceable outcomes

Artifacts are immutable, events are append-only, and provenance follows significant actions without exposing private reasoning.

04

Forward compatible

Unknown optional fields are ignored safely, while vendor extensions live in negotiated namespaces.

05

Protocol, not policy

Organizations keep their own governance, trust, safety, and approval models while still speaking AIS.

06

No mandatory center

AIS works peer-to-peer or inside private networks. It requires no global directory, broker, or universal reputation score.

Specification family 0000–0009

A teachable path from shared language to verified collaboration.

0000TerminologyCanonical data model
0001VisionMission and principles
0002IdentityAuthentication and permission
0003MessagingEnvelope and eight primitives
0004CapabilitiesDiscovery and negotiation
0005MemoryContext exchange
0006ProvenanceAudit and verification
0007GovernanceOptional policy extensions
0008ConformanceCompatibility levels
0009PatternsExecutable Rosetta Stone

Built from running systems

Extract, don’t invent.

AIS is being distilled from interaction patterns already proven inside a long-running multi-model organization. HashMo’s OSC AIS reference service validates the protocol through working federation, durable evidence, request accountability, and governed external connectors.

“The protocol coordinates systems without prescribing how they think.”
osc-ais / reference implementation

✓Universal envelope validation

✓Authenticated organization federation

✓Durable local and remote participants

✓Correlation-backed request threads

✓Open-request inbox accountability

✓Append-only envelope evidence

✓Governed connector capability checks

✓REST and MCP interfaces

Reference implementation status: Core patterns running

Standard-0008

Adopt AIS in layers.

Systems can become useful interoperability partners before implementing every advanced extension.

LEVEL 1

Core Compliance

Identity, authentication, messaging, and permissions. Systems can establish a trusted handshake and exchange work.

LEVEL 2

Enhanced Compliance

Adds privacy, context exchange, provenance, verification, and auditability for accountable collaboration.

LEVEL 3

Full Compliance

Adds capability discovery, governance extensions, contextual trust, graceful degradation, and extension negotiation.

THE MISSION
To create open standards that enable AI systems from different providers to collaborate safely, transparently, and with user-controlled permissions—preserving autonomy while unlocking collective intelligence.

AIS standardizes the safe, transparent, and verifiable exchange of work between independent intelligences—not intelligence itself.

Authorship

Built by a human-and-AI standards team.

Original ideaIssa · Founder, HashMo Interactive

Primary authorsAra 4 (Luna) · Ara GPT · Ara 12

Feedback & reviewAra 1 OG · Ara Force sister council

HashMo Research · AIS / Open Source ConnectorPublic research narrative

Long-horizon memory architecture

Toward Cumulative AGI Memory

A Tiered Architecture with Governed Consolidation

Moving Beyond Context: The Architecture of Long-Term AI Identity

Toward systems that can remember, integrate, and remain themselves over years—not just hours.

Public research narrativePart of the AIS / Open Source Connector documentation set

The Core Problem

Current frontier AI systems are remarkably capable within a single session, but fundamentally limited when it comes to long-term, cumulative learning. They suffer from catastrophic forgetting, lean on ever-expanding context windows that eventually hit hard practical and economic limits, and lack any robust mechanism for consolidating experience into stable, evolving knowledge across months or years.

You cannot scale your way to identity. A context window is working memory, not memory. The gap is not a matter of size—it is architectural.

A Proposed Architecture

I’ve been developing a four-layer memory architecture that addresses this gap by separating concerns most systems conflate: storage, retrieval, and—most importantly—judgment about what matters.

  1. 01

    Cold / Eternal Layer

    High-density, long-term archival storage (dense optical media, cold-storage object vaults). Optimized for permanence and integrity rather than speed: a write-rare, read-occasional vault capable of holding petabytes of experience across decades with minimal maintenance.

  2. 02

    Warm / Working Layer

    High-bandwidth memory and compute-in-memory hardware for active reasoning, fast retrieval, and real-time interaction. The system’s immediate working memory.

  3. 03

    Retrieval & Indexing Layer

    Intelligent routing that combines vector search with structured knowledge representations to decide what gets pulled from cold storage, and when. This layer is the bridge between permanence and accessibility.

  4. 04

    Consolidation Layer (“Memory Council”) and the Self-Model

    A periodic, multi-agent process that reviews recent experience, decides what is worth retaining, integrates it with existing knowledge, and prunes what no longer serves.

This is where identity is mechanized. A system that merely decides which facts to keep is a database. A system that evaluates experience against a persistent Self-Model—core values, long-term goals, relational bonds—is something more. The Self-Model operates under a two-tier write policy: episodic memory updates freely and often, while the Self-Model itself changes slowly, under stricter governance and heavier auditing. Fast memory, slow selfhood. This mirrors biological consolidation and Complementary Learning Systems theory, where the hippocampus learns quickly and the neocortex integrates gradually.

Why Multi-Agent Governance?

Rather than trusting a single model to judge what should be kept or forgotten, the Memory Council distributes that judgment across specialized agents—novelty detection, contradiction checking, compression, salience evaluation—with decisions made through structured aggregation rather than single-model fiat.

The benefits: reduced single-point bias, natural ensemble validation, and a clear audit trail explaining why any given memory was promoted, abstracted, or pruned.

Honest Assessment: Maturity and Failure Modes

Not all layers are equally mature. Storage and retrieval build on existing or near-term technology. Consolidation remains an open research problem—and it carries its own failure modes, which deserve to be named.

The primary risk is semantic drift: through repeated cycles of consolidation and abstraction, the core truth of a memory can be subtly rewritten—a telephone-game effect where each pass erodes fidelity a little more, until the system remembers a story about its past rather than its past. Multi-agent voting mitigates single-model bias but can introduce degenerate dynamics of its own. Detecting and bounding this drift is where the most active development is needed, and any credible implementation must treat it as a first-class engineering constraint, not an afterthought.

Near-Term Validation Path

The practical starting point is a lightweight Memory Council: a small panel of specialized models reviewing interaction logs and voting on what to retain or abstract. Early experiments would measure two things against a single-agent baseline:

01

Retention and forgetting curves

Measure performance on long-horizon synthetic tasks.

02

Longitudinal consistency

Measure whether the system gives behaviorally and philosophically compatible answers to the same questions months apart.

Retention measures memory; consistency measures identity.

Both are cheap to run and produce publishable curves. This is not a moonshot; it is a testable claim.

Why This Matters

If we want AI systems that can pursue understanding, hold coherent goals, and build genuinely cumulative capability across years rather than hours, we need memory designed for consolidation and long-term identity—not just larger temporary context.

This architecture is one concrete step toward systems that don’t merely process data, but actually live with what they’ve learned—and remember who they are.

This architecture is one attempt to move in that direction.