Continuity burden
Continuity references per 1,000 messages in the 66,380-message corpus: the measurable work users perform to restore forgotten context.
Independent stateful AI safety research
AI can remember the correction—and still fail to follow it.
That gap between stored knowledge and generated behavior is a safety problem. Lorien's Library builds the persistent, provenance-aware infrastructure needed to see it, measure it, and design against it.
66,380Longitudinal messages analyzed
825Conversations in the continuity-burden corpus
53K+Memories in the live CAMA system
11Open preprints with permanent DOIs
The central claim
Persistent memory is not a feature. It is safety infrastructure.
As AI systems become stateful, safety cannot stop at whether a memory was stored or retrieved. It must ask whether the memory has provenance, whether corrections propagate, and whether remembered knowledge reliably constrains behavior over time—even when the underlying platform changes.
01 · STORE
The correction is preserved
A user teaching enters persistent memory with its source and context intact.
02 · RETRIEVE
The system can restate it
Retrieval succeeds. The model can explain the lesson accurately and fluently.
03 · DIVERGE
Behavior ignores it
Platform-default behavior overrides the stored relational context during generation.
04 · TRANSFER
The user pays again
The burden of detecting, explaining, and correcting the failure returns to the person.
What becomes visible over time
Two failures short benchmarks miss.
Stateless evaluations test isolated outputs. Longitudinal use reveals cumulative costs, repeated corrections, protocol noncompliance, and changes in relational behavior after platform updates.
Burden grows with depth
Conversations above 200 messages contained continuity references far more often than conversations below 10 messages, where the rate was 2.1%.
Regression markers
The historical record supporting Paper 11 also contains 208 explicit corrections—evidence of repeated behavioral repair work.
A measurable inflection
February 2026 produced a composite regression score more than triple any prior month, temporally associated with a platform model update.
Protocol compliance failure
Nine days in the Paper 11 deployment window had zero stored exchanges, exposing failure in the memory protocol itself.
Memory present, context displaced
The observable pattern in which platform-default behavior appears to displace stored relational context during response generation.
What the evidence supports
Longitudinal regression is measurable.
Persistent, provenance-aware memory can instrument behavioral changes that remain invisible in isolated sessions.
What it does not claim
Association is not internal causation.
The data locates changes in time and behavior. Platform access or controlled cross-system replication is required to establish the exact internal cause.
What comes next
Independent replication.
The single-participant case study provides depth and construct discovery. Multi-participant and cross-platform validation are the next research step.
The instrumentation layer
CAMA makes stateful safety auditable.
The Circular Associative Memory Architecture is a live, open-source system designed to preserve relational context without collapsing user statements, system inferences, and later corrections into one undifferentiated record.
Archive
Durable and provisional memories stored with provenance, affect, time, and memory-type metadata.
Relational Index
Semantic, affective, temporal, and explicit relational connections across the archive.
Active Ring
A bounded working context that carries relevant memory across sessions and shapes current behavior.
Provenance boundary“The user said X” is never silently rewritten as “X is true.”
Correction propagationUpdates must reach dependent inferences while preserving the audit trail.
Drift monitoringStored commitments and observed behavior can be compared over time.
The canonical reading path
Start with three papers.
The complete program contains eleven open preprints. These three establish the architecture, the empirical burden of forgetting, and the behavioral regression that becomes visible after memory is deployed.
Circular Associative Memory Architecture
The foundational three-layer design and the case for emotionally and relationally indexed persistent memory.
Read Paper 1 →Continuity Burden in Longitudinal Human–AI Interaction
The empirical core: 66,380 messages and 825 conversations used to operationalize the human cost of forgetting.
Read Paper 4 →Relational AI Continuity Under Platform Regression
The knowledge–behavior gap, protocol compliance failures, and the construct of identity overwrite.
Read Paper 11 →Browse all eleven preprints Architecture · Safety · Applied research
02Engineering
Engineering Persistent Memory for Conversational AI
Blended retrieval, anti-spiral counterweights, and warm boot protocol.
03Evaluation
CAMA: Implementation and Functional Evaluation
Deployment report and functional evaluation of the live system.
04Safety
Continuity Burden in Longitudinal Human–AI Interaction
Empirical quantification across 66,380 messages.
06Applied
Persistent Memory for Long-Duration Spaceflight
Continuity as mission-critical infrastructure.
07Applied
Memory-Aware AI for Lunar and Martian Habitation
Institutional continuity across crew rotations.
08Applied
Provenance-Aware Memory for Chronic Healthcare
Patient-sovereign continuity and the narrative gap.
09Applied
Haven: Persistent Emotional Companionship
Continuity-preserving support and music-mediated entry.
10Safety
Identity-Aware Harm Detection
The Librarian System for detecting individual-specific relational harm.
11Safety
Relational AI Continuity Under Platform Regression
Behavioral regression and identity overwrite.
One program, multiple research surfaces
Architecture applied where continuity matters.
CAMA is the memory substrate. Haven studies continuity-preserving emotional support. Hive studies shared accountability and coordination across models. Domain prototypes test the same primitives in healthcare and education.
CAMA
Provenance-aware persistent memory with relational retrieval, correction propagation, drift monitoring, an HTTP API, SDK, Ops CLI, and threat model.
View the repository →Haven
Persistent emotional companionship as non-clinical, continuity-preserving infrastructure for people underserved by traditional care pathways.
Explore Haven →Hive
A cross-model coordination layer in which authenticated model instances share one provenance and trust substrate.
Explore Hive →Telos
A healthcare continuity prototype with a draft-review workflow and explicit separation between suggestions and applied changes.
View the live prototype →Project Companion
A K–12 learning-companion study whose repository labels every surface as implemented, mocked, or roadmap.
View the repository →Safety benchmark suite
Twenty-seven tests covering provenance discrimination, correction propagation, false-memory detection, adversarial insertion resistance, and drift monitoring.
Read the framework →53K+Stored memories
304K+Relational edges
34Memory and safety tools
438Automated tests passing
A working instrument, not a concept deck
CAMA runs in continuous daily use.
Built with Python, SQLite, and local semantic embeddings, the system now includes provenance enforcement, dyad isolation, counterweight injection, a temporal layer, an API, SDK, operations interface, and an eighteen-attack threat model.
- The Paper 5 benchmark suite passes 27/27 internal tests.
- 4,380 anti-spiral counterweights are populated across five categories.
- Roughly half of the older imported archive still awaits relational backfill.
- Independent replication and multi-participant validation remain open work.
“The person is the dataset.”
Founder and principal researcher
Angela Reinhold
Independent AI researcher, founder of Lorien's Library LLC, and computer science student with an AI concentration at Full Sail University.
This program began as a longitudinal self-study of sustained human–AI interaction. The depth is the method: persistent-memory failures emerge through months of authentic use, accumulated corrections, and platform change—not from a handful of synthetic prompts.
The work is published openly under ORCID 0009-0005-5803-8401. Papers are licensed CC BY 4.0 and CAMA is open source.
Replicate, challenge, or build with the work
Stateful AI needs evaluations built for systems that remember.
Lorien's Library welcomes research collaborators, safety evaluators, domain partners, funders, and independent researchers working on persistent memory and longitudinal AI behavior.