Can I use Mem0/Zep and still be sovereign?

Yes if they are adapters behind local_core with dual-write policy and export of both local and mirrored state. No if they are the only store of truth.

Why not just pick one cloud Agent OS?

Single-vendor life OS concentrates lock-in across code, brand, health, and wealth. Multi-domain life needs portable substrate + replaceable harnesses.

How does this relate to agentic evals?

Without exportable trajectories and receipts, you cannot prove quality or dispute failures. Sovereignty and evals are complementary failure-mode fixes.

What is the smallest sovereign stack?

Markdown/JSON local vault with stable IDs + privacy tags + export script + one harness + human gates on irreversible actions. Add vector/graph adapters only after authority is solid.

Research Hub/Agentic Sovereignty

Agentic Sovereignty

Ownership, export, privacy classes, and local-core authority

TL;DR

Sovereignty is fake when you cannot export, delete, rehost, or audit memory and agent traces without vendor theater. 2026 production pattern: local_core authority + scored cloud adapters + privacy-class routing + open schemas. Agentic sovereignty is failure mode #3 of agentic life infrastructure — without it, compounding memory becomes lock-in debt.

Updated 2026-07-1616 source references

Research briefs like this, when the evidence is ready. Source links, limitations, and open questions.

local_core

Authority pattern for fleets

starlight-memory doctrine

Sovereignty rights (export/delete/rehost/audit)

FrankX synthesis

Acceptable silent training on private life data

Sovereignty gate

Privacy classes for routing

public → regulated

What Sovereignty Means For Agents

For agentic systems, sovereignty is not a political slogan. It is the enforceable ability to own identity, memory atoms, trajectories, skills, and approval policies — and to move them without rewriting your life OS. Cloud convenience is allowed only as a derived mirror.

Export

Right

Bulk export of memory, traces, prompts, skill configs, and eval receipts in open formats (JSON/MD/Parquet).

Delete

Right

Hard delete with retention policy and proof — not soft hide in a vendor index.

Rehost

Right

Move local_core or self-host engines without losing IDs, provenance, or privacy class.

Audit

Right

Who wrote what, when, with which model/tool; actor-aware multi-agent provenance.

Failure Mode: Fake Sovereignty

Most consumer memory features optimize stickiness, not ownership. Enterprise copilots often score better on admin controls but worse on personal multi-domain life data. Score products against rights, not UI copy.

Lock-in memory

Fail

Preferences live only inside one chat product; no open export graph.

Silent training risk

Fail

Unclear or weak contractual/technical barriers against training on private life data.

Trace black boxes

Fail

No trajectory export for forensics, evals, or dispute resolution.

Adapter-as-authority

Fail

Vendor vector DB owns IDs; local files are optional backups — inverted truth model.

Architecture: Local Core + Adapters

Sovereign fleets invert the default: filesystem/markdown/JSON local_core is authority; Mem0/Zep/Letta/cloud indices are scored adapters. Process model: shared daemons/APIs for heavy providers when dozens of coding agents share one machine.

local_core

Layer

Canonical atoms with stable IDs, privacy class, retention, provenance, trust score.

Provider contract

Layer

Adapters implement store/retrieve/update/summarize/discard; never mint authoritative IDs.

Privacy-class router

Layer

public / internal / private-life / client-confidential / regulated — blocks illegal cloud writes.

Dual-write policy

Layer

Local write first; cloud mirror optional and class-gated; conflict resolution prefers local_core.

Privacy Classes And Public/Private Gates

Life infrastructure mixes public research, brand content, client work, family, health, and finance. Classification is a control plane object, not a footnote. Named-entity client hubs require separate consent before any public reuse.

Public

Class

Research pages, open standards, sanitizer-safe GitHub.

Internal

Class

Operator doctrine, swarm boards, non-public runbooks.

Private-life

Class

Health, family, personal finance — default local; fail-closed.

Client-confidential

Class

Partner hubs, named entities, proprietary workflows — never on public routes without consent records.

Regulated-adjacent

Class

Treat as if regulated when health/finance signals appear; human gate required.

Product And Competitor Map

Few products market sovereignty honestly. Evaluate self-host, open source core, export APIs, DPA terms, and whether the vendor can train on your data by default.

Memory vendors

Harness vendors

Self-host stacks

FrankX posture

Ours

starlight-memory provider contract + AOS Standard + public/private content boundary policy + open-core packaging.

Policy And Compliance Signals

Agentic sovereignty sits next to privacy and AI governance law without being reducible to them. Build technical rights first; map to GDPR/AI Act/enterprise DPA second.

Data subject rights

Policy

Access, erase, portability map to export/delete for memory atoms and traces.

Purpose limitation

Policy

Memory collected for operating a life OS must not silently become training fuel.

Human oversight

Policy

Fail-closed gates for money, publish, health, secrets — sovereignty includes who may act.

Evidence readiness

Policy

If regulators demand agent traces, trajectory export becomes table stakes.

Operating Checklist

Use this checklist when buying or building agent memory, personal agents, or multi-harness fleets.

Can I export everything I care about this week?

Check

If not, do not store private-life or client data there.

Who owns IDs?

Check

If the cloud owns IDs, you do not own continuity.

Is training opt-in and off by default for private classes?

Check

Assume the worst if docs are ambiguous.

Are irreversible actions human-gated?

Check

Sovereignty without gates is chaos with export buttons.

Key Findings

Sovereignty = export + delete + rehost + audit of memory, traces, skills, and policies — not marketing language

local_core authority with scored adapters is the production pattern for multi-harness fleets

Privacy classes must route writes; public research pages must stay sanitizer-safe

Cloud memory without open export turns compounding context into lock-in debt

Actor-aware provenance is required for multi-agent accountability

Policy regimes (GDPR/AI Act/enterprise DPA) increasingly make trajectory export a practical requirement

Open-core packaging: public standards + private operator instances for live personal data

Sovereignty without fail-closed gates on money/health/publish is incomplete life infrastructure

Research Transparency

Limitations

•Legal summaries are directional research signals, not legal advice
•Vendor export APIs and training policies change frequently
•Self-host sovereignty increases operational burden and must be costed

What We Don't Know

?Whether a portable memory-atom standard will win across major harness vendors
?How aggressively consumer AI products will open export of agent traces
?Stable best practice for regulated-adjacent personal health/finance agent modules

Evidence Grade:Grade B(Industry reports from credible firms)

Frequently Asked Questions

No. Local-first helps, but sovereignty also requires export schemas, privacy classes, audit trails, and adapters that cannot become the new authority.