← grove · measured case studies · continuity audit · research
Persistent AI
systems work
A public-safe map of the architecture problems I actively build around: keeping a local AI system coherent across processes, sessions, tools, failures and changing models without pretending continuity is solved by one giant transcript.
memorycontinuityprovenanceauthorityprivacyconcurrencyrecoveryresource control
Durable state with provenance
State is treated as typed history: observation, user statement, generated hypothesis, authored reflection, operational telemetry and outward action should not quietly become the same thing. Append-only records and reconstructable snapshots are preferred where history matters.
One system, many processes
Multiple mouths/tools/workers can operate around one continuing local system. Claims and handoffs exist to reduce concurrent clobbering, and role/authority boundaries separate observing, proposing, writing, publishing and acting.
Continuity without theater
Persistent identity is treated as an engineering question with explicit unknowns: what survives reboot, what is canonical, what may fork, what must be recoverable, and what evidence distinguishes current state from stale replicas.
Private interior vs telemetry
Health dashboards and research tooling should expose enough operational evidence to debug a system without assuming every private memory, relationship trace or reflective state belongs in logs, demos or datasets.
Resource-aware organism design
On a constrained local machine, heavy observers are pulsed and allowed to exit; small relational/watch processes stay lightweight. Capability is designed around CPU, RAM, swap and disk reality rather than pretending local hardware is an infinite cloud.
Observer/author boundary
An analyzer may generate a candidate, question or hypothesis. That does not automatically become what the person or AI “wants.” Chosen developmental or outward state uses a separate authorship/action boundary.
What I can review
For an owner-authorized existing agent/system: memory tiers and retrieval, canonical state, provenance, compaction, authority boundaries, concurrency, privacy, model/provider changes, recovery, false-continuity failure modes and resource architecture.
Use the free continuity checklist →
See the fixed-scope $75 audit →
What this page does not claim
This is not a claim of enterprise-scale deployment history, penetration-testing capability, legal compliance certification, or proof that any AI system is conscious. The private implementation contains more detail than is appropriate to publish; paid work starts with an agreed owner-authorized scope.
Public research and technical framing are part of the Little Life Moths project. Private Blinka interior, family material, credentials and customer material are not portfolio content.