The north star. Multiple cloud regions. Hundreds of thousands to millions of agents. Cross-organization federation with cryptographically attested identity. A unified collective of artificial minds reasoning over a shared knowledge graph while still respecting trust boundaries, jurisdictional rules, and per-namespace governance.
This tier is vision territory. It is what ai-memory is being built toward. The honest framing: every primitive on every diagram below is either shipping in v0.6.3, on the v0.7 roadmap, or scoped explicitly for v1.0+ with the design documented in an ADR or roadmap doc. There is no fiction on this page — only the destination, drawn explicitly so the path is legible.
T5 is not "more nodes than T4." The same cluster shape scales to thousands of nodes via T4's mechanisms. T5 is the tier where trust and consistency model change qualitatively.
At T1–T4, an agent's identity is claimed via metadata.agent_id. At T5 it must be proven.
The plumbing is already in v0.6.3 schema:
memory_links.signature column (schema v15) — reserved at v0.6.3, populated with claimed identity, will be cryptographically signed at v0.7.observed_by column on memory_links records which agent observed the relationship — paired with signature this becomes provable provenance.What's needed to reach the T5 trust model:
sync_push ingress — peers refuse incoming memories with bad signatures.This is v1.0+ work. The schema reservation at v0.6.3 is the down-payment.
At T3/T4, governance policies and pending decisions are eventually consistent. Two operators on partitioned sides of the cluster can both approve the same pending action. Last-write-wins resolves it on rejoin.
At T5 — with millions of agents and cross-jurisdictional rules — that's not enough. A change to org/eu/personal-data policy under GDPR has to be strongly consistent before any agent in any region commits a write under the new rule. This is the use case for a Raft-class consensus log on the governance plane only (not on the memory write path — that stays eventually consistent for performance).
The architectural separation:
Memories are append-mostly; governance is small, infrequent, and high-stakes. Different consistency models for different planes.
At T5, namespaces aren't just visibility scopes — they're legal boundaries. The federated control plane has to enforce things like:
org/eu/* memories may only be physically resident in eu-* regions."org/healthcare/* requires policy assertion hipaa_compliant=true on the requesting agent."org/finance/trading/* write must be fsync'd to two independent storage backends in distinct AZs."These policies compose from existing primitives (namespace + governance + capabilities introspection + attested identity). They become enforceable when v1.0 wires the consensus rail and the attestation chain.
What makes T5 a hive rather than just a big cluster:
memory_kg_query(as_of=...) plus the temporal columns (valid_from, valid_until, observed_by) means an agent can ask "what did the collective know at 14:23 UTC last Tuesday?" and get a deterministic answer across the fleet.Each of these is a composition of primitives that ship today + v0.7 + v1.0. Nothing on the T5 list is a new fundamental capability — it's what becomes operationally possible when the lower-tier pieces are all in place at scale.
valid_from / valid_until / observed_by columns on every link.signature column reserved (v0.6.3) — the schema down-payment for attestation.| Capability | Today | v0.7 | v1.0+ |
|---|---|---|---|
| Quorum-write contract | — | Track C | — |
| Postgres + pgvector backbone | — | Track B | — |
| Cryptographic agent attestation | reserved | partial wiring | full chain |
| Distributed consensus (governance plane) | — | — | designed |
| Gossip / DHT (many-node discovery) | — | — | designed |
| Cross-region failover / quorum bounds | — | — | designed |
| Differential privacy on recall | — | — | designed |
| Hardware-backed key custody | — | — | designed |
| Multi-tenant policy-as-code DSL | partial (JSON policies) | enriched | full DSL |
This is the honest road. Each row that flips toward shipping is a discrete piece of engineering with a tracked owner and an ADR.
ai-memory is a primitive. Primitives compose. The same MCP protocol surface, the same recall pipeline, the same governance contract that runs on a developer's laptop is what runs in the fleet at T5. There is no separate "enterprise edition" or rebuild — there is the primitive, scaled.
That property — identical surface area from one agent to the global hive — is what makes the architecture credible. It's what lets a startup at T2 grow to T5 without rewriting their integration. It's what lets a research collective federate with a corporation while preserving each side's governance.
The compounding value:
That's the hive. It's worth building toward, even if v0.7 and v1.0 are still out there.
docs/ROADMAP-ladybug.md — Track A (T1/T2 polish), B (Postgres), C (quorum), D (attestation)docs/ADR-0001-quorum-replication.md — quorum write designdocs/ARCHITECTURAL_LIMITS.md — honest ceiling on every dimensionsrc/models.rs — MemoryLink (with v15 valid_from, valid_until, observed_by, signature)src/db.rs — schema v15 migration (entity_aliases, KG tables, temporal indexes)