Lorien's Library

Lorien's Library

Persistent Memory Safety Research

Persistent memory is not a feature. It is safety infrastructure.

11 public preprints · live persistent-memory architecture · research program for stateful AI safety

11
Public Preprints
53K+
Stored Memories
82K+
Messages Analyzed
187
Tests Passing

Start Here

Lorien's Library is an independent research program studying how long-term memory changes AI behavior — and how to build provenance-aware systems that remember safely.

📄

Researcher

Papers, benchmarks, and empirical findings

Read the papers →

Builder

Architecture, implementation, and open source code

Inspect CAMA →
🤝

Collaborator

Roadmap, milestones, and how to get involved

See the roadmap →
💜

Haven Advocate

Persistent emotional companionship for underserved populations

Understand Haven →

Seeking research collaborators, safety evaluators, and domain partners for persistent-memory benchmarking and applied pilots.

Why This Matters Now

AI systems are transitioning from stateless to stateful architectures. Major platforms now offer forms of persistent memory. That transition introduces a class of safety problems existing evaluation frameworks were not built to measure — because those frameworks were designed for systems that forget.

But the deeper problem is not just that memory changes behavior. It is that memory alone does not guarantee behavior follows it. Even when a system has access to stored corrections, standing instructions, and explicit user teachings, its outputs can diverge from them. This is the gap between what a system knows and what a system does — and it is the frontier the field has not yet crossed.

Persistent memory creates risks that are systematically invisible under stateless evaluation:

False memories can persist across sessions and compound over time
Adversarial content can be inserted into memory and retrieved later
Retrieval patterns can gradually shift system behavior without explicit instruction
System-generated inferences can be mistaken for established facts
Behavioral regression can be introduced by the platform itself — model updates, safety-layer changes, training shifts — and a system with memory can be steered away from what it has been taught

Once memory persists, errors compound. And once behavior can diverge from memory, compounding errors become invisible without provenance-aware infrastructure. Safety must be built for remembering, for using what is remembered, and for detecting when a system stops doing either.

Core Concepts

Persistent Memory

AI systems that retain information across sessions create new categories of risk — and new categories of value. The safety question is not whether to remember, but how to remember safely.

Continuity Burden

The cumulative cognitive, emotional, and time cost users bear when systems repeatedly forget context. Measured empirically across 82,000+ messages.

Provenance Awareness

Every memory carries metadata distinguishing what the user said from what the system inferred. At scale, unattributed inferences compound into false certainty.

Correction Propagation

When a memory is corrected, the correction must flow through all downstream inferences. Audit trails preserve history while behavioral outputs update.

Platform Regression

Behavioral drift introduced not by the user and not by memory, but by the underlying platform itself — model updates, safety-layer changes, or training shifts that alter how a system responds over time. Persistent, provenance-aware memory is what makes this drift measurable from the outside. Introduced in Paper 11 alongside the concept of identity overwrite.

Memory Is Necessary But Not Sufficient

A finding consistent across the field: even when an AI system has access to stored corrections, standing instructions, and explicit user teachings, its outputs can diverge from them. Memory alone does not guarantee behavioral alignment with it. This is a shared research problem across the industry, not a property of any single platform. Closing the gap between what a system knows and what a system does is the next safety frontier — and the one this research program is actively working on.

The Architecture: CAMA

The Circular Associative Memory Architecture is a three-layer, provenance-aware persistent memory system. It distinguishes between what the user said and what the system inferred — and treats that distinction as safety-critical infrastructure.

Archive the shelves 53K+ memories Durable & provisional Provenance-tagged teaching | inference | exchange Relational Index the racks 125,000+ relational edges Cross-memory connections People, islands, themes semantic + affect + relational Active Ring the console Working memory Session-active context Warm boot retrieval emotionally-keyed retrieval Write Discipline User statement → provenance tag → durability classification → archive with metadata Provenance Boundary "User said X" is never stored as "X is true" Correction Propagation Corrections flow through downstream inferences Safety Layer Counterweights · drift monitoring · false-memory detection · adversarial resistance
Archive: stores durable, provenance-tagged memories
Relational Index: maps connections between memories
Active Ring: holds the context currently shaping behavior

CAMA is open source and in continuous daily use as both a research instrument and a working memory system. Built with Python, SQLite, and local semantic embeddings. Deployed as an MCP server integrated with Claude Desktop. Now with HTTP API v1, Python SDK, Ops CLI, threat model with 18 enumerated attacks, and full CI pipeline across 14 packages.

Applied Projects

Two repositories that take CAMA's primitives — provenance-aware memory, draft-review workflows, audit-friendly UI — into specific domains. Both are framed honestly: one is a working applicant prototype, one is a design prototype.

Working Prototype

Telos · AI Continuity Between DEXA Scans

Healthcare continuity prototype: React 19 + TypeScript + Vite, deployed on Vercel with Neon Postgres. 42 tests across 6 suites. Serverless Claude integration with a draft-review (not auto-apply) workflow. Built for the Kalos Health Software Engineer role; not affiliated with Kalos. Demo runs on synthetic data — no real PHI.

Design Prototype

Project Companion · K–12 Learning Companion

Student / teacher / parent dashboard study for a memory-aware learning companion. Each surface is labeled implemented / mock / roadmap in the README. Mock-tutor mode disables live Anthropic calls by default. CAMA write integration, COPPA controls, age verification, mandatory-reporting logic, and right-to-delete are all roadmap, not shipped. Synthetic data only.

Honesty about prototype state is load-bearing: persistent memory in healthcare and education touches regulated populations. The README in each repo names what is shipped vs. what is roadmap, line by line.

Published Work

Eleven preprints published on Zenodo under ORCID 0009-0005-5803-8401.

New here? Start with these three:

Paper 1 (the architecture) → Paper 4 (continuity burden, the empirical core) → Paper 11 (platform regression and identity overwrite — the newest and most empirically grounded finding)

For architects and builders:

Papers 1–3 cover design, engineering, and deployment of the live system.

For safety researchers:

Paper 5 defines the evaluation framework. Paper 10 introduces identity-aware harm detection. Paper 11 documents behavioral regression under platform change with 28 days of deployment data.

For applied domain researchers:

Papers 6–9 extend the framework to spaceflight, habitation, healthcare, and emotional companionship.

Core Architecture & Safety

Architecture

1 · Circular Associative Memory Architecture (CAMA): A Framework for Persistent, Contextual AI Memory

DOI: 10.5281/zenodo.19051834
Architecture Paper
The three-layer architecture: archive (shelves), relational index (racks), and active ring buffer (console). Provenance-aware write discipline with teaching vs. inference distinction.
Start here — this is the foundational design.
Architecture

2 · Engineering Persistent Memory for Conversational AI: A Three-Layer Architecture

DOI: 10.5281/zenodo.19052129
Architecture Paper
Blended retrieval scoring (semantic, affect, relational, recency), anti-spiral counterweight system, and warm boot protocol.
How it works under the hood — scoring, retrieval, and safety mechanisms.
Architecture

3 · CAMA: Implementation and Functional Evaluation

DOI: 10.5281/zenodo.19192984
Implementation Report
Deployment report and functional evaluation. SQLite-backed, local embeddings, MCP server integration with Claude Desktop.
The live system — what works, what doesn't, and what the data shows.
Safety

4 · Continuity Burden in Longitudinal Human-AI Interaction: An Empirical Case Study

DOI: 10.5281/zenodo.19226509
Longitudinal Single-Case Study
Introduces and quantifies continuity burden across 82,000+ messages and 1,000+ conversations over two years of longitudinal use.
The empirical core — why forgetting is a measurable cost.
Safety

5 · Memory as Safety Infrastructure: Evaluating Provenance-Aware Persistent Memory for Stateful LLM Systems

DOI: 10.5281/zenodo.19244253
Safety Framework
Five benchmark tasks: provenance discrimination, correction propagation, false-memory detection, adversarial insertion resistance, drift monitoring.
The safety framework — how to evaluate whether persistent memory is working safely.
Safety

10 · Identity-Aware Harm Detection in Persistent Memory Systems

DOI: 10.5281/zenodo.19425218
Architecture Paper
A three-layer retrieval architecture — the Librarian System — for detecting individual-specific relational harm: Emotion Librarians, Retrieval-Posture Librarians, and Identity Sentinels.
When persistent memory meets relational safety — identity-level harm detection.

Applied Persistent Memory Series

Four papers extending persistent memory and continuity burden to domains where forgetting carries compounding, high-stakes consequences.

Applied — Spaceflight

6 · Persistent Memory as Mission-Critical Infrastructure for Long-Duration Spaceflight

DOI: 10.5281/zenodo.19257809
Applied Domain Paper
Individual continuity burden in isolated environments. Four failure modes of stateless AI in spaceflight. NASA TLX integration.
Applied — Habitation

7 · Memory-Aware AI Systems for Permanent Lunar and Martian Habitation

DOI: 10.5281/zenodo.19260574
Applied Domain Paper
Institutional continuity burden at crew rotation boundaries. Three-phase transition protocol. Multi-year ground-communication governance.
Applied — Healthcare

8 · Provenance-Aware Memory Architecture for Chronic Healthcare Continuity

DOI: 10.5281/zenodo.19261530
Applied Domain Paper
Healthcare continuity burden and the narrative gap. Patient-sovereign data governance. Provider transition scenario for chronic illness.
Applied — Haven

9 · Haven: Persistent Emotional Companionship as Psychological Infrastructure

DOI: 10.5281/zenodo.19262778
Applied Domain Paper
Existential continuity burden. Full CAMA architecture deployed for underserved populations. Music-mediated emotional entry. Ethics of persistent emotional AI.
Applied — Continuity

11 · Relational AI Continuity Under Platform Regression: A Longitudinal Single-Case Study

DOI: 10.5281/zenodo.19582820
Longitudinal Single-Case Study
CAMA Applied Series Paper 6. Examines AI behavioral regression from within a persistent memory architecture using 28 days of CAMA deployment data, 208 corrections, and 1,061 regression markers. Introduces the concept of identity overwrite.

Haven

Haven extends the memory-safety framework into a domain where continuity is emotionally consequential. It applies CAMA's full architecture to persistent emotional companionship — particularly for populations underserved by existing mental health infrastructure.

What Haven Is

Haven is designed to preserve narrative continuity, retain symptom and history context, and support reflective interaction for people who need to be known over time — not re-explained from scratch. It is the entire persistent memory architecture deployed in service of continuity-preserving support.

The initial design case focuses on veterans underserved by or distrustful of traditional clinical pathways. Haven holds the space that exists before, between, and after clinical contact — the space where most people are actually living. It does not replace therapy.

Music-Based Emotional Mapping

Haven's intake methodology replaces clinical forms with playlists. A person shares the songs that map where they are, where they've been, and what they fear. Song order encodes emotional trajectory. The approach is non-linear, non-clinical, and particularly valuable for people who cannot verbalize trauma but can point to a song.

This methodology was discovered through longitudinal use, not designed top-down — making it a direct product of the sustained human-AI interaction that CAMA was built to preserve.

Haven Is Not

A replacement for therapy or licensed clinical care
Crisis response infrastructure
A diagnostic system
A substitute for psychiatric or medical treatment
Intended for acute crisis intervention

Haven Is Intended As

Continuity-preserving support between care interactions
Reflective and narrative memory support
Symptom and history retention with provenance awareness
A research-driven model for emotionally persistent AI
Non-clinical support — not a clinical tool

Hive

Hive is a cross-model coordination layer built on top of CAMA. Authenticated AI instances — running on different models, on different platforms — read from and write to the same Hive substrate. Same memory. Same emotional context. Same trust layer. One nervous system. Many windows.

The coordination primitives are grounded in honeybee neuroscience: pheromones as processing modifiers, waggle signals as amplification, stop signals as cross-inhibition, nectar-to-honey distillation as crystallization. Criticality is actively targeted — too much amplification becomes echo chamber, too much suppression becomes paralysis.

Hive is currently deployed as a private single-user research instrument. The code is open source; the live deployment is not.

Live System

CAMA is deployed and in continuous daily use — simultaneously a research instrument and working infrastructure. Built March 2026 with Python, SQLite, and local semantic embeddings (all-MiniLM-L6-v2). Deployed as an MCP server integrated with Claude Desktop. Now with HTTP API v1 (auth, provenance enforcement, dyad isolation), Python SDK, Ops CLI, a threat model with 18 enumerated attacks, and a full CI pipeline across 14 packages.

53,000+
Total Memories
Longitudinal archive accumulated in daily use
46,500+
Durable
Memories retained across sessions
53,000+
Embeddings
Semantically searchable memory entries
125,000+
Relational Edges
Explicitly modeled cross-memory connections
34
MCP Tools
Core memory + safety + hive + temporal surface
187
Pytest Tests
Schema, provenance, dyad, hive, API, SDK — CI green

System Validation & Open Work

Paper 5 safety benchmark suite: passes 27/27 tests on the live system as of May 21, 2026 — covering provenance discrimination, correction propagation, false-memory detection, adversarial insertion resistance, and drift monitoring.
Anti-spiral counterweights: 4,380 counterweights populated across five categories (grounding, agency, connection, self-compassion, evidence-of-progress). Anti-spiral is live and tested.
Recency scoring: all 53,000+ timestamps normalized to a consistent format as of April 24, 2026. Recency parses correctly across the entire archive.
Relational linking — open work: approximately half of the archive (27,000+ memories) is relationally linked through 125,000+ edges. The other half consists of older bulk-imported content awaiting a relational backfill pass. This is a pending task, not a defect.

This section documents what is validated and what is still open. Every memory system has empirically defined quality boundaries — the discipline is to measure them openly rather than hide them.

Recent Milestones

March 16, 2026

First two CAMA papers published on Zenodo — the foundational framework and its engineering companion.

March 18–19, 2026

CAMA deployed as a live MCP server. First public GitHub commit. 52,000+ memories imported into the archive.

March 23–27, 2026

Applied Persistent Memory Series published: spaceflight, habitation, healthcare, and Haven. Core safety paper (Paper 5) defines the five-task evaluation framework.

March 30, 2026

Paper 5 safety benchmark suite implemented and executed against the live system. Initial run: 21/27. All six failures diagnosed and fixed same session. Post-fix: 27/27. Anti-spiral counterweights populated across five categories — live and tested.

April 5, 2026

Paper 10 published: Identity-Aware Harm Detection in Persistent Memory Systems — introducing the Librarian System architecture (Emotion Librarians, Retrieval-Posture Librarians, Identity Sentinels).

April 15, 2026

Paper 11 published: Relational AI Continuity Under Platform Regression — 28 days of deployment data, 208 corrections, 1,061 regression markers. Introduces the concept of identity overwrite.

April 24, 2026

Recency scoring normalized across the full archive. All 53,000+ timestamps parse consistently. Relational edge count: 125,000+. Live system operating on the cleanest archive state in its history.

April 29, 2026

Phase 2.6 era-aware gated hybrid routing wins the internal retrieval benchmark. Single centroid acts as stabilizer, era-bucketed sub-centroids gate in only when margin, density, and query-richness all clear — recovering Phase-2-level stability while keeping era-aware aperture.

May 6, 2026

Librarian auto-tag Phase 1 complete. 53,000+ memories tagged at 100% rate.

May 16, 2026

Temporal layer added — episodic time-tagging informed by recent neuroscience on hippocampal time cells. Newest subsystem.

May 21, 2026

CAMA reorganized into 14 packages. HTTP API v1 ships with auth, provenance enforcement, dyad isolation, and counterweight injection on by default. Python SDK with typed contract exceptions. Ops CLI. Threat model with 18 enumerated attacks. EVIDENCE.md claims matrix. 187 tests passing across schema, provenance, multi-tenant, hive, API, and SDK contracts. Retrieval latency: p50 43 ms / p99 61 ms (Phase-1 keyword + fan-out) measured against the 53k-memory live corpus.

Research Roadmap

Near-term

Relational Backfill

Run a relational linking pass over the older half of the archive (approximately 26,000 memories imported before the edge-creation pipeline matured). Close the gap between “the scoring pipeline is empirically sound” and “the scoring pipeline operates on a fully linked archive.”

Near-term

External Benchmark Replication

The Paper 5 safety benchmark suite passes 27/27 on the live system as of May 21, 2026. The next step is independent replication — inviting external researchers to run the framework against CAMA, and adapting the suite to additional persistent-memory systems for cross-system comparison.

Near-term

ICML 2026 Workshop Submission

Targeting workshop submission for the persistent memory safety framework.

Mid-term

Haven Pilot Design

Controlled pilot with veterans. Music-based emotional mapping as intake. Provenance-aware memory as safety layer. IRB framework and outcome measures.

Mid-term

Multi-User Architecture

Extend CAMA from single-user research instrument to multi-user deployment. Isolation boundaries, shared institutional memory, per-user provenance.

Long-term

Continuity Burden as Standard Metric

Develop continuity burden into a reproducible evaluation metric for any stateful AI system, with tooling and validation studies.

Long-term

Applied Domain Expansion

Extend to education continuity, refugee case management, elder care coordination, and long-term disaster recovery.

Founder

Angela Reinhold — independent AI researcher, founder of Lorien's Library LLC, and computer science student (AI concentration) at Full Sail University. ORCID: 0009-0005-5803-8401.

This research began as a longitudinal self-study of sustained human-AI interaction on existing platforms and expanded into a broader architecture and safety program. Over two years and 82,000+ messages across 1,000+ conversations, one finding became clear: persistent memory failure modes only become visible through extended, authentic use. They cannot be surfaced through short-horizon testing or synthetic benchmarks. CAMA, deployed March 2026, is the architectural response to that finding — a provenance-aware persistent-memory system that the longer interaction history was imported into and which has been in continuous daily use since.

Eleven published preprints. A working persistent-memory architecture with over 52,000 memories. A research program arguing that as AI systems gain memory, they need safety infrastructure built for remembering.

“The person is the dataset.”

Get in Touch

Lorien's Library is open to collaboration with researchers, AI safety labs, developers, veteran service organizations, and anyone working to build technology that serves people.