Future substrate report — what donto is still missing

Living document. 2026-06-12. A century-scale substrate gap report: what donto already has, what is still missing, and what has to become true before AI systems can treat donto-like memory as durable infrastructure. For the full reports index, see Reports.

donto now has a credible kernel. It is not yet a civilization substrate.

The kernel is real: append-only claims, bitemporality, paraconsistency, evidence links, policy, lenses, standing, loss reports, release artifacts, and agent discovery. The missing work is larger than another feature wave. donto needs the identity, source, semantic, cryptographic, operational, governance, economic, and human machinery that lets memory survive many models, many institutions, and many generations.

TL;DR

• The current win: donto has the right memory discipline: claims are evidence-linked, contradiction-tolerant, time-aware, lens-scoped, and exported with loss instead of pretending that every view is complete.
• The current gap: the system is still mostly a strong local kernel. It does not yet have universal identity, deep source packages, federation, cryptographic proof chains, governance protocols, scale architecture, or mature human review surfaces.
• The real target: not one world database. The target is an interoperable protocol where many stores can exchange claims, evidence, policies, contradictions, standing, and loss without flattening the context.
• The next move: make the kernel boring, then federate it, then govern it. A memory substrate only matters if outsiders can verify it, communities can constrain it, and future systems can migrate it.

donto should not become "the memory all AI queries." It should become the discipline by which AI memory stays attributable, contestable, policy-aware, and reconstructable.

Century-substrate scorecard

1. What donto already has

The completed execution work turned many implicit conventions into durable substrate behavior:

That is a serious kernel. The rest of this report is the gap between a kernel and infrastructure.

2. The target shape

A century-scale AI memory substrate has to satisfy several properties at once:

1. It remembers without pretending memory is truth.
2. It preserves disagreement without turning disagreement into chaos.
3. It lets many agents contribute without giving any agent total authority.
4. It supports private, community-governed, public, commercial, and scientific knowledge under different rules.
5. It lets future agents reconstruct why a belief existed, which evidence supported it, which policies constrained it, and what later work changed it.
6. It survives software rewrites, hardware generations, organization collapse, legal changes, and model-family turnover.
7. It makes extraction, inference, alignment, ranking, and forgetting accountable.

The shape is therefore a protocol, not a single database:

source bytes
  -> anchored observations
  -> claims + time + policy
  -> contradiction-aware lenses
  -> standing + loss reports
  -> signed releases / query results
  -> agents that know what they may use, quote, train on, or publish

3. Missing foundation map

Identity, sources, and semantics

Why this is first: memory fails fastest at the boundary between "same thing", "possibly same thing", "not same thing", and "same only under this lens." If identity and predicates drift silently, a century of memory becomes a bag of strings with timestamps.

Governance, trust, and lifecycle

Hard part: not "delete or keep." The hard part is recording what can be forgotten, what must remain auditable, what can survive only as a hash/policy event, and what cannot be used for a purpose even if it is still known.

Federation, APIs, and scale

Federation rule: copy claims only with their evidence, policy, lineage, standing context, contradictions, and loss reports. Anything less is not interop; it is a misleading export.

Multimodal, causal, and model-native memory

An AI using donto should know when it is reading evidence, when it is reading a summary, when it is reading a contested claim, and when it is not authorized to know something.

Human, economic, adversarial, and formal layers

Formal methods should not try to prove every stored claim. They should prove the substrate does not lie about its own behavior.

4. What the future user experience should feel like

An agent encountering a claim should be able to ask:

That is accountable memory: not omniscience, but a reconstructable trail from answer back to source, policy, disagreement, and loss.

5. Roadmap by horizon

6. Core risk

The biggest risk is not that donto lacks features. The biggest risk is that it becomes a powerful local system without becoming an interoperable public discipline.

A private memory substrate can be useful. A century-scale AI substrate must be legible to outsiders, hostile auditors, future maintainers, communities whose knowledge is stored, agents that did not exist when the data was written, and institutions that need to rely on it after the original operators are gone.

7. Bottom line

donto is missing the layers that turn a strong knowledge kernel into durable civil infrastructure:

The execution plan made donto credible as a governed paraconsistent substrate. The next challenge is to make it durable, federated, inspectable, and worth trusting for generations.

See also: reports index · the shape of donto's return · single-shot vs agentic memory · benchmarks.