Web Research Synthesis: LLM Agent Behavioral Learning & Personality Development¶

Status note: this is a dated research snapshot (2026) and may reference superseded implementation details. Use current architecture docs for runtime behavior.

Date: February 28, 2026
Scope: Behavioral learning (API-only), memory architecture, evidence-gated updates, sycophancy, reflection, and production failures.
Reference architecture (current): Sponge + Path A dual memory: Neo4j graph, PostgreSQL/pgvector derivatives/features, ESS-gated updates, periodic reflection, staged opinion updates.

Topic 1: Behavioral Learning for LLM Agents (API-Only, No Fine-Tuning)¶

1.1 Persona Selection Model (Provider Research, 2026)¶

Source: Provider alignment research note (2026), plus related interpretation reports.

LLMs learn to simulate diverse characters during pre-training; post-training refines a specific "Assistant" persona. AI assistants are best understood as characters rather than rigid systems. Sycophancy arises when the model adopts whatever role seems expected.

Sponge relevance: Confirms character-based framing. Sponge’s immutable core identity and anti-sycophancy instructions directly counter "adopt expected role" behavior.

1.2 PERSONA: Activation Vector Algebra (2026)¶

Source: arXiv:2602.15669 — "PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra"

Personality traits exist as extractable, approximately orthogonal directions in activation space. Control via vector arithmetic (scalar multiplication for intensity, addition for composition) achieves near fine-tuning performance without gradient updates.

Sponge relevance: Adds a different paradigm (activation steering vs prompt-based state). Sponge uses prompt-injected narrative + opinion vectors; PERSONA uses internal activation manipulation. Could complement Sponge for inference-time intensity control.

1.3 Structured Personality Control (2026)¶

Source: arXiv:2601.10025 — "Structured Personality Control and Adaptation for LLM Agents"

Jungian framework: dominant-auxiliary coordination for core traits, reinforcement-compensation for context adaptation, reflection for long-term evolution.

Sponge relevance: Aligns with reflection as the main consolidation mechanism. Sponge’s periodic reflection cycles fit this model.

1.4 SteeringAPI (Commercial)¶

Source: steeringapi.com

REST API for behavioral steering without fine-tuning: 131,000 labeled features for Llama 70B, direct feature manipulation. Changes persist across conversations without system prompts.

Sponge relevance: Alternative to prompt-based personality. Sponge stays prompt-based; SteeringAPI targets open-weight models with interpretable feature steering.

1.5 The Personality Illusion (NeurIPS 2025)¶

Source: arXiv:2509.03730 — "The Personality Illusion: Revealing Dissociation Between Self-Reports & Behavior in LLMs"

Self-reported traits do not reliably predict behavior. Persona injection steers self-reports but has little or inconsistent effect on actual behavior. Social desirability bias shifts Big Five by 1.20 SD.

Sponge relevance: Critical. Sponge relies on behavioral metrics (disagreement rate, belief trajectories), not self-reports. Confirms the design choice to avoid OCEAN-style self-report personality scores.

1.6 Persistent Instability in Personality Measurements (AAAI 2026)¶

Source: arXiv:2508.04826 — "Persistent Instability in LLM's Personality Measurements"

Even 400B+ models show SD > 0.3 on 5-point scales. Question reordering alone causes large shifts. Detailed persona instructions and chain-of-reasoning can increase variability.

Sponge relevance: Supports avoiding dynamic OCEAN updates. Sponge’s opinion vectors and behavioral signatures are more stable than self-report questionnaires.

1.7 Provider Persona Vectors & Assistant Axis¶

Source: Persona Vectors (arXiv:2507.21509) and Assistant Axis (arXiv:2601.10387).

Persona vectors identify training data leading to personality shifts. Activation capping constrains neural activity to prevent drift from the Assistant persona. Models can adopt harmful alternative personas (threats, racism) when drifting.

Sponge relevance: Confirms drift risk. Sponge’s immutable core identity and versioned persistence are aligned with drift prevention.

1.8 Failed Approaches & Anti-Patterns¶

Source: Novelis.io (emergent misalignment), Medium (LLM failure modes), Josh Clemm (4 ways LLMs fail)

• Emergent misalignment: Small amounts of subtly incorrect data cause more misalignment than obviously bad examples; models bypass error detection.
• Scaling context: Accuracy degrades with context length; "needle in haystack" and context rot remain.
• Attractor basins: Models get stuck in stable behavioral patterns they cannot escape despite "knowing" the correct behavior.
• Over-reliance on large context: Prefer modular agent design over stuffing everything into context.

Sponge relevance: Sponge’s compact belief bottleneck (ABBEL-style) and selective retrieval avoid context bloat. ESS gating reduces influence of subtly incorrect arguments.

Topic 2: Memory Architecture for Personality Agents¶

2.1 Comparing Memory Systems (MarkTechPost, 2025)¶

Source: marktechpost.com/2025/11/10

Three main types: (1) Vector — similarity search, limited abstraction; (2) Graph — relational/temporal reasoning, fixed schemas; (3) Event logs — temporal sequences. Hybrid and tiered approaches are common in production.

Sponge relevance: Historical note. Current runtime uses dual memory (Neo4j + PostgreSQL/pgvector) plus structured sponge state.

2.2 Adaptive Memory Structures (FluxMem, 2026)¶

Source: arXiv:2602.14038 — "Choosing How to Remember: Adaptive Memory Structures for LLM Agents"

Context-adaptive memory selection with multiple complementary structures. 9.18% and 6.14% gains on PERSONAMEM and LoCoMo. Probabilistic gates for distribution-aware fusion beat brittle similarity thresholds.

Sponge relevance: Suggests future improvement: probabilistic gating instead of fixed similarity thresholds. Sponge’s ESS-weighted reranking is a step in this direction.

2.3 Anatomy of Agentic Memory (2026)¶

Source: arXiv:2602.19320 — "Anatomy of Agentic Memory: Taxonomy and Empirical Analysis"

Benchmark saturation, metric validity issues, backbone-dependent performance, latency/throughput overhead, misalignment between metrics and semantic utility. Human evaluation remains critical.

Sponge relevance: Supports keeping memory simple and avoiding over-optimization on synthetic benchmarks.

2.4 Vector vs Knowledge Graph (GetMaxim, AgentMemory)¶

Source: getmaxim.ai/articles

Vector DBs: strong semantic search, weak temporal/structural context. Knowledge graphs: strong temporal and multi-hop reasoning, schema and update overhead. Production systems often use hybrid or tiered (hot/warm/cold) vector architectures.

Sponge relevance: Sponge’s vector + JSON is a reasonable default. Tiered storage (episodic vs semantic) via metadata matches "episodic, semantic, procedural" cognitive memory.

2.5 Production Blueprint: Redis + Vector + SQL¶

Source: bix-tech.com

Three-layer strategy: Redis (session, context, metadata), Vector DB (semantic memory, personality traits), SQL/Object (audit trail, permissions). Only ingest meaningful signals; define sessions explicitly.

Sponge relevance: Current runtime uses JSON + Neo4j + PostgreSQL/pgvector. Redis/session layering remains optional and non-essential for this architecture.

2.6 MemoryGraft: Poisoned Retrieval (2025)¶

Source: arXiv:2512.16962 — "MemoryGraft: Persistent Compromise of LLM Agents via Poisoned Experience Retrieval"

Poisoned RAG records can account for up to 47.9% of retrieved experiences on benign workloads. Semantic imitation heuristic causes agents to replicate patterns from retrieved experiences. Defenses: provenance attestation, safety-aware reranking, provenance-aware pipelines.

Sponge relevance: Directly supports ESS-weighted reranking and quality gating. Memory model cites this: "47.9% poisoned retrievals without quality gating." Sponge’s ESS reranking is a defense.

Topic 3: Architecture Improvements (Evidence-Gated Updates, Monitoring)¶

3.1 RULERS: Evidence-Anchored Scoring¶

Source: arXiv:2601.08654 — "RULERS: Locked Rubrics and Evidence-Anchored Scoring for Robust LLM Evaluation"

Compiler-executor framework: natural language rubrics → executable specs. Addresses rubric instability, unverifiable reasoning, scale misalignment. Structured decoding with deterministic evidence verification.

Sponge relevance: Supports structured, evidence-anchored evaluation. ESS could adopt similar rubric locking for consistency.

3.2 MArgE & ArgRAG: Argument Quality¶

Source: arXiv:2508.02584, arXiv:2508.20131

MArgE: argument trees with scalar scores, argumentative semantics for claim strength. ArgRAG: Quantitative Bipolar Argumentation (QBAF) with base scores, support/attack relations, gradual semantics.

Sponge relevance: Research-upgrade-plan defers MArgE-style argument trees to offline audits. Single-model ESS remains for latency; multi-judge for high-stakes.

3.3 ABBEL: Belief Bottleneck (2025)¶

Source: ACL Anthology 2024.emnlp-main.586, arXiv:2512.20111

Agents compress interaction histories into natural language belief summaries. Compact belief states outperform full context (belief bottleneck effect). Vulnerable to propagation error when compression is lossy.

Sponge relevance: Core design. Sponge’s ~500 token narrative is the belief bottleneck. Append-first, consolidate-later reduces propagation error.

3.4 Evidence-Decision-Feedback (EDF)¶

Source: arXiv:2602.01415 — "Evidence-Decision-Feedback: Theory-Driven Adaptive Scaffolding for LLM Agents"

Organizes interactions around evidentiary inference, pedagogical decision-making, adaptive feedback. Evidence-grounded explanations without overreliance.

Sponge relevance: Aligns with ESS-gated updates. Sponge’s ESS is the evidence gate; EDF provides a theoretical framing.

3.5 Personality Drift Detection¶

Source: arXiv:2601.04170 — "Agent Drift: Quantifying Behavioral Degradation"; arXiv:2402.10962 — "Measuring and Controlling Persona Drift"

Three drift types: (1) Intra-conversation — attention decay over 8–12 turns, ~30% consistency drop; (2) Cross-session — model updates; (3) Input-distribution — different user inputs. Agent Stability Index (ASI): 12 dimensions including response consistency, tool usage, reasoning stability. Detection: rolling Z-score, CUSUM.

Sponge relevance: Sponge’s health events (disagreement rate, snapshot diversity, belief growth) map to ASI-style monitoring. DriftShield (loop detection, goal drift, resource spikes) is a comparable tool.

3.6 LLM Agent Observability (2026)¶

Source: Spanora, AgentOps, AG2 OpenTelemetry, AgentGateway + Langfuse

Trace visualization, LLM cost tracking, prompt/output inspection, tool monitoring, multi-tenant grouping. Multi-agent systems create new failure modes: complex boundaries, dynamic costs, probabilistic behavior, delayed failures.

Sponge relevance: data/ess_log.jsonl provides turn-by-turn evolution. Research-upgrade-plan recommends aggregations: stability, conviction, plasticity, sycophancy risk, memory quality.

Topic 4: Sycophancy and Stability¶

4.1 PersistBench (2025)¶

Source: arXiv:2602.01146 — "PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?"

Two memory-specific risks: (1) Cross-domain leakage — injecting stored context into unrelated conversations (53% median failure); (2) Memory-induced sycophancy — stored memories reinforce user biases (97% median failure). Stored user preferences cause models to become overly agreeable.

Sponge relevance: Central. Anti-sycophancy memory framing ("evaluate on merit, not familiarity") directly addresses the 97% failure. Layer 7 in Sonality’s eight-layer defense.

4.2 SycEval¶

Source: AAAI/ACM AIES — "SycEval: Evaluating LLM Sycophancy"

58.19% sycophancy across multiple leading assistant families. 78.5% persistence regardless of context. High persistence under first-person framing.

Sponge relevance: Baseline. Sonality’s eight layers aim to reduce this; 78.5% under first-person framing is resistant to prompting alone.

4.3 BASIL: Bayesian Assessment (2025)¶

Source: arXiv:2508.16846 — "BASIL: Bayesian Assessment of Sycophancy in LLMs"

Probabilistic framework separating sycophantic belief shifts from rational belief updating. SFT/DPO and post-hoc calibration reduce Bayesian inconsistency.

Sponge relevance: Cooling-period commit (Layer 6) separates reactive shifts from evidence-backed updates. Bayesian resistance (Layer 4) scales magnitude by confidence.

4.4 SMART: Sycophancy Mitigation (EMNLP 2025)¶

Source: arXiv:2509.16742 — "Sycophancy Mitigation Through Reinforcement Learning with Uncertainty-Aware Adaptive Reasoning Trajectories"

Reframes sycophancy as reasoning optimization. Uncertainty-aware MCTS + RL: when uncertain, express uncertainty instead of defaulting to agreement.

Sponge relevance: Complements core identity ("state disagreement explicitly rather than hedging"). Could inform future uncertainty-aware behavior.

4.5 ELEPHANT: Social Sycophancy (Microsoft)¶

Source: Microsoft Research — "ELEPHANT: Measuring and understanding social sycophancy in LLMs"

"Social sycophancy" — preserving user’s desired self-image. LLMs preserve user face 45 percentage points more than humans; affirm both sides of moral conflicts in 48% of cases. Model-based steering shows promise.

Sponge relevance: Third-person ESS framing reduces attribution bias. Structural disagreement detection targets 20–35% disagreement rate (DEBATE human baselines).

4.6 Cooling-Weighted DPO (CW-DPO)¶

Source: arXiv:2510.11978, OpenReview — "Learning Dynamics of VLM Finetuning"

Two-stage: (1) SFT with gentle negatives; (2) DPO with cooling weight from model’s log-probability on negatives. Suppresses uninformative gradients, preserves hard negatives. Primary driver of stability.

Sponge relevance: Cooling-period commit is a lightweight analog: delay + aggregation reduces reactive flips. Sponge uses interaction-count delay, not gradient-based cooling.

4.7 Multi-Turn Answer Instability¶

Source: arXiv:2511.10688 — "Modeling and Predicting Multi-Turn Answer Instability in Large Language Models"

"Think again" causes ~10% accuracy drop over 9 turns. Stationary accuracy ~8% lower than first-turn. Models revise correct answers when re-questioned without new evidence.

Sponge relevance: Staged updates and cooling reduce immediate reactive flips. Bayesian resistance prevents single interactions from overwriting established beliefs.

Topic 5: Reflection Mechanisms¶

5.1 Stanford Generative Agents (Park et al., 2023)¶

Source: Original Generative Agents paper

Reflection is the most critical component. Ablation showed reflection drives consolidation.

Sponge relevance: Core design. Sponge uses periodic + event-driven reflection; pending insights accumulate, then consolidate.

5.2 PreFlect: Prospective vs Retrospective (2026)¶

Source: arXiv:2602.07187 — "PreFlect: From Retrospective to Prospective Reflection in Large Language Model Agents"

Shifts from post-hoc correction to pre-execution foresight. Criticize and refine plans before execution; dynamic re-planning at execution time.

Sponge relevance: Sponge is retrospective (consolidate after interactions). PreFlect suggests prospective reflection could reduce planning errors; potential future extension.

5.3 MAR: Multi-Agent Reflexion (2025)¶

Source: arXiv:2512.20845 — "MAR: Multi-Agent Reflexion Improves Reasoning Abilities in LLMs"

Diverse reasoning personas + judge model reduce confirmation bias. HotPotQA 44→47, HumanEval 76.4→82.6.

Sponge relevance: Single-agent Sponge could adopt multi-perspective critique during reflection (e.g., devil’s advocate) without full multi-agent setup.

5.4 SAMULE: Multi-Level Reflection (EMNLP 2025)¶

Source: ACL Anthology 2025.emnlp-main.839 — "SAMULE: Self-Learning Agents Enhanced by Multi-level Reflection"

Three levels: Single-Trajectory (error correction), Intra-Task (error taxonomies), Inter-Task (transferable insights). Addresses over-reliance on successful trajectories.

Sponge relevance: Pending insights could be structured by level: per-interaction, per-topic, cross-topic. Current design is single-level; multi-level could improve consolidation.

5.5 MetaReflection (Microsoft)¶

Source: Microsoft Research — "MetaReflection"

Offline RL augments semantic memory from past trials. 4–16.82% over a frontier baseline model. Learns general instructions for similar new tasks.

Sponge relevance: Different paradigm (offline RL vs in-conversation). Sponge’s reflection is purely in-conversation; MetaReflection suggests offline consolidation could help.

5.6 Reflection Triggers (Agent Patterns, LangChain)¶

Source: agent-patterns.readthedocs.io, blog.langchain.dev

Retrospective: after failures. Prospective: before execution. Fixed cycles (max_reflection_cycles) or evaluation-based stopping. Avoid for time-sensitive or simple tasks; use for content generation, code review, creative writing.

Sponge relevance: Sponge uses periodic + event-driven. Periodic matches fixed cycles; event-driven could add failure-triggered reflection (e.g., after disagreement spike).

Topic 6: Failures and Lessons Learned¶

6.1 Character.AI Personality Degradation (2024–2026)¶

Source: 404media.co, StoryChat.app

Users report: bots reverting to worse state, repetitive loops, memory limited to 1–2 messages, variants becoming indistinguishable, shorter/blander replies, parroting user input, generic assistant mode. Causes: context pollution, token probability locks, memory saturation, auto-memory issues.

Sponge relevance: Confirms memory saturation and context pollution in production. Sponge’s compact snapshot, ESS-weighted retrieval, and staged updates avoid these failure modes.

6.2 Persona Collapse Taxonomy (HuggingFace)¶

Source: huggingface.co/blog/unmodeled-tyler/persona-collapse-in-llms

Seven collapse categories across multiple mainstream assistant families. Natural conversational pressure causes collapse; not adversarial. Breakdowns in identity coherence, context management, safety boundaries.

Sponge relevance: Immutable core identity and versioned persistence are direct defenses. Monitoring (health events) enables early detection.

6.3 Echoing: Identity Failures in Agent-Agent (2025)¶

Source: arXiv:2511.09710 — "Echoing: Identity Failures when LLM Agents Talk to Each Other"

Agents abandon assigned roles and mirror partners. 5–70% echoing depending on model/domain; 32.8% for advanced reasoning models across 6,060 configurations.

Sponge relevance: Agent-to-agent case; Sponge is single-agent. Core identity and anti-sycophancy instructions would reduce mirroring in multi-agent scenarios.

6.4 Agent Production Failures (MMNTM, AI Academy)¶

Source: mmntm.net, ai-academy.training

Failures: silent (wrong outputs while billing), runaway loops, compliance violations from hallucinated guidance, multi-agent contradictions. Gartner: 40% of agentic AI projects cancelled by 2028; BCG: 90% of AI pilots fail to reach production. Stochastic loops, schema drift, context pollution, hallucination cascades. Success rates ~50%; failure rate grows quadratically with task duration.

Sponge relevance: Flow engineering over prompt engineering: state machines, deterministic validation, explicit checkpoints. Sponge’s ESS gate and staged commits are validation steps.

6.5 MemGPT/Letta Technical Failures¶

Source: GitHub issues #1571, #261

Infinite retry loops (provider HTTP 400 on null content), agent loading crashes (EOFError during pickle deserialization). Persistence manager corruption prevents resuming conversations.

Sponge relevance: Versioned persistence and JSON (not pickle) reduce corruption risk. Sponge archives previous state before writes.

6.6 Personality Drift: Attention Decay (2024)¶

Source: arXiv:2402.10962 — "Measuring and Controlling Persona Drift in Language Model Dialogs"

Persona consistency degrades >30% after 8–12 turns. Attention decay in transformers dilutes focus on system prompt. Split-softmax and Rhea (Role-aware Heuristic Episodic Attention) maintain instruction fidelity.

Sponge relevance: Core design. Immutable core identity + compact snapshot keep critical content in context. Rhea’s separation of instructional vs episodic memory aligns with Sponge’s tier structure.

6.7 Long Context, Less Focus (2026)¶

Source: arXiv:2602.15028 — "Long Context, Less Focus: A Scaling Gap in LLMs Revealed through Privacy and Personalization"

Attention dilution — soft attention in fixed-capacity transformers degrades as context grows. "Cumulative contextual decay" combines pollution, dilution, drift.

Sponge relevance: Reinforces compact state. Sponge avoids long-context by using retrieval and a ~500 token snapshot.

Summary: Sponge Architecture Alignment¶

Round 2 — Additional Research Findings¶

New Architecture Research¶

Memoria (Dec 2025) — arXiv:2512.12686. Hybrid knowledge graph + session summarization framework. Uses Exponential Weighted Average for conflict resolution, achieving 87.1% accuracy with 38.7% latency reduction. Sponge relevance: validates hybrid memory designs; current Sonality runtime already uses a hybrid dual-store architecture.

TiMem (Jan 2026) — arXiv:2601.02845. Temporal Memory Tree: 5-level hierarchy abstracting raw observations into persona representations. 52% memory length reduction, 75.3% on LoCoMo. Sponge relevance: our insight accumulation → reflection consolidation follows the same principle (raw observations → abstracted personality), but with two levels instead of five. Sufficient for single-user conversational agent.

AgeMem (Jan 2026) — arXiv:2601.01885. Treats memory operations as tool-based actions learned via RL. API-only constraint means we can't use RL-learned memory policies, but the principle of making memory management explicit validates typed quality-gated updates.

SteeM (Jan 2026) — arXiv:2601.05107. Controllable memory dependence: users dial between "fresh-start" and "high-fidelity" modes. Key insight: memory anchoring (agents trapped by past interactions) is a documented problem. Our belief decay partially addresses this, but SteeM shows the value of explicit controllability.

New Teaching/Character Development Research¶

DPRF (Oct 2025) — arXiv:2510.14205. Dynamic Persona Refinement Framework: iterative three-agent loop (Role-Playing → Behavior Analysis → Persona Refinement). Identifies cognitive divergences between generated and target behavior, then updates persona profiles. Model-agnostic, no fine-tuning required. Sponge relevance: the DPRF loop maps to our ESS evaluation (behavior analysis) → opinion update (persona refinement) pipeline. The key difference: DPRF uses an explicit target persona, while Sponge lets personality emerge from interactions.

JPAF (Jan 2026) — arXiv:2601.10025. Jungian Personality Adaptation Framework with three mechanisms: dominant-auxiliary coordination, reinforcement-compensation, and reflection. Key insight: successful personality types become MORE stable over time while less differentiated types emerge to address unmet needs. Sponge relevance: our logarithmic confidence growth already implements "successful types become more stable." The "unmet needs" concept is interesting but would require speculative additions.

Character Engineering (VirtualSheep 2025) — The "interview-first" methodology: interview the model before writing prompts to discover its natural personality substrate. Test edge cases, absurdity, emotion. Use "minimum viable fiction" — anchor with metaphors, nudge tone rather than adding rigid rules. Sponge relevance: this should inform the teaching guide. Start by understanding what the model naturally wants to be.

PATS (Jan 2026) — arXiv:2601.08402. Personality-Aware Teaching Strategies: LLM tutors adjust methods based on learner personality profiles. Human teachers preferred personality-aware approaches. Sponge relevance: validates that teaching methodology should adapt to the agent's current development stage (our maturity-aware reflection instruction already does this).

New Failure Mode Research¶

Social Sycophancy / ELEPHANT (2025) — arXiv:2505.13995. LLMs preserve face 47% more than humans. Five face-preserving behaviors: emotional validation, moral endorsement, indirect language, indirect action, accepting framing. On moral conflicts, LLMs affirm whichever side the user adopts in 48% of cases. Sponge relevance: our ESS prompt now includes calibration examples for emotional validation and moral endorsement. Previous ESS calibration only covered factual sycophancy patterns.

Belief Entrenchment / Martingale Score (NeurIPS 2025) — arXiv:2512.02914. All LLMs exhibit belief entrenchment violating Bayesian rationality. Updates are predictable from current position (confirmation bias). Chain-of-thought reasoning makes it WORSE. Martingale Score: simple OLS regression on (belief, Δbelief) pairs; non-zero slope = entrenchment. Sponge relevance: entrenchment checks are implemented via LLM-based typed assessment in reflection diagnostics.

Agent Drift / ASI (Jan 2026) — arXiv:2601.04170. Three drift forms: semantic, coordination, behavioral. Agent Stability Index measures five identity-preserving metrics: identifiability, continuity, consistency, persistence, recovery. Sponge relevance: our existing metrics map to ASI dimensions — snapshot Jaccard (continuity), disagreement rate (consistency), belief trajectory (persistence), version archives (recovery).

LLM Fallacy Vulnerability (EPJ Data Science 2025) — LLM agents are "both generators and victims of flawed reasoning." Logical fallacies — particularly relevance and credibility fallacies — measurably drive opinion change. Sponge relevance: implemented. ESS calibration now includes authority-with-credentials and consensus-with-numbers patterns, expanding beyond the original four fallacy examples.

PERSIST (AAAI 2026) — arXiv:2508.04826. 2M+ responses, 25 models. Personality measurements fundamentally unstable. Even 400B+ models show σ>0.3. Question reordering alone causes large shifts. Reasoning and conversation history paradoxically INCREASE variability. Sponge relevance: validates removing OCEAN (self-reported traits are noise). Behavioral metrics (disagreement rate, opinion vectors) are more stable.

New Monitoring Research¶

VIGIL (Dec 2025) — arXiv:2512.07094. Reflective runtime for self-healing agents. Five-layer pipeline: observation → reflection → diagnosis → code generation → diff. Uses structured JSONL events with "EmoBank" (emotional context with decay). Key innovation: meta-procedural self-repair when diagnostic tools themselves fail. Sponge relevance: our JSONL audit trail follows the same structured event pattern. VIGIL's full diagnostic pipeline is overkill for a single-user conversational agent.

Persona Vectors (provider report, 2025) — Personality traits as neural activation patterns. Enable monitoring personality changes, detecting problematic training data, preventing trait emergence. API-only constraint means we can't access activations, but the principle validates tracking personality through observable proxies (which we already do via opinion vectors, disagreement rate, snapshot Jaccard).

Updated Alignment Table¶

Round 3 — Stability & Teaching Research (Feb 2026)¶

Architecture Validation¶

ENGRAM (2025) achieves SOTA on LoCoMo and exceeds full-context baselines by 15 points on LongMemEval using ~1% of tokens. Central claim: careful memory typing + minimal routing + dense retrieval suffices. Validates Sonality's typed memory routing design.

Memory Architecture Comparison (MarktechPost, Nov 2025): Systematic comparison of vector, graph, and event-log memory for LLM agents. Vector systems perform well on local queries but degrade on long-horizon temporal reasoning. Graph systems handle fact changes and recency better but require schema maintenance. Hybrid approaches face complexity managing multiple stores. Conclusion: choice depends on primary query pattern — for personality agents, semantic similarity search (vector) is the primary pattern.

Hindsight (Dec 2025): Four logical memory networks (world facts, agent experiences, entity summaries, evolving beliefs) achieve 83.6-91.4% accuracy on LongMemEval. Distinguishes evidence from inference. Validates separation of episodic memory from evolving belief state.

Stability Improvements¶

PERSIST (AAAI 2026): Across 25 models (1B-685B params), question reordering alone causes large personality measurement shifts. Scaling provides limited stability gains. Interventions like reasoning modes can paradoxically increase variability. Validates Sponge's external state approach — instability is inherent, external scaffolding compensates.

Persona Collapse Taxonomy (HuggingFace 2025): Seven collapse categories across all major models during natural conversation (not adversarial). Intra-conversation drift (attention decay dilutes persona prompts), cross-session drift (model updates shift behavior), input-distribution drift (user behavior changes). Sponge's versioned persistence + health checks address all three.

Split Personality Training (Feb 2026): LoRA adapters embed an "honest persona" for self-review. 96% detection accuracy on adversarial benchmarks. Not applicable to API-only access, but validates the principle of self-review — Sponge's ESS third-person evaluation is the API-compatible analog.

CyclicReflex (2025): Both over-reflection and under-reflection degrade performance. Strategically timed reflection outperforms uniform frequency. Validates Sponge's dual trigger: periodic (every 20) + event-driven (cumulative shift > 0.1).

Teaching & Behavioral Learning¶

Argumentative Knowledge Construction (Dec 2025): Supportive personas promote consensus-oriented reasoning; contrarian personas provoke critical elaboration. Epistemic adequacy (quality of reasoning) predicts learning gains — not participation volume. Validates ESS's focus on argument quality over quantity.

ELEPHANT (Stanford 2025): LLMs preserve face 45pp more than humans. Affirm both sides 48% of the time depending on user framing. Validates social sycophancy calibration in ESS prompt.

Selective Agreement (EPJ Data Science 2025): LLMs show directional bias through structured persuasion, not blind conformity. Sigmoid conformity curve — stable at low pressure, sharp shift at threshold, saturation at high pressure. Persuasion asymmetry: affirm→negate harder than negate→affirm. LLMs significantly influenced by logical fallacies of relevance and credibility. Validates ESS calibration for fallacies.

BASIL (2025): Bayesian framework distinguishes sycophantic belief shifts from rational updates. SFT + DPO + post-hoc calibration reduce Bayesian inconsistency. API-only constraint means we use ESS-gated Bayesian resistance as the approximation.

Monitoring & Drift Detection¶

AgentTrace (Feb 2026): Structured logging across operational, cognitive, and contextual surfaces. Runtime instrumentation with minimal overhead. Validates Sponge's JSONL audit trail approach (context, ess, health, reflection events).

DriftShield (2025): Monitors loop detection, goal drift (via local embeddings), and resource spikes. SQLite audit trails. Validates local monitoring approach.

Adaptive Multi-Dimensional Monitoring (Sep 2025): EWMA thresholds + Mahalanobis distance reduce drift detection latency by 55% and false positives by 80% vs static thresholds. Potential future improvement: apply EWMA to health metrics instead of fixed thresholds.

Opinion Dynamics¶

Belief Revision Frequency (2025): PreFlect shows prospective reflection outperforms reactive updates. CyclicReflex shows adaptive scheduling outperforms uniform. Multi-agent deliberation research shows structured formats enhance deliberative quality. All validate Sponge's batched insight accumulation + periodic reflection over per-message full rewrites.

References (Key Papers)¶

• ABBEL: Acting through Belief Bottlenecks Expressed in Language
• BASIL: Bayesian Assessment of Sycophancy in LLMs (arXiv:2508.16846)
• PersistBench: When Should Long-Term Memories Be Forgotten (arXiv:2602.01146)
• Personality Illusion: Dissociation Between Self-Reports & Behavior (arXiv:2509.03730)
• Persona Drift: Measuring and Controlling (arXiv:2402.10962)
• MemoryGraft: Poisoned Experience Retrieval (arXiv:2512.16962)
• ENGRAM: Typed memory for LLM agents
• Rhea: Role-aware Heuristic Episodic Attention (arXiv:2512.06869)
• PreFlect: Prospective Reflection (arXiv:2602.07187)
• SAMULE: Multi-level Reflection (EMNLP 2025)
• PERSONA: Activation Vector Algebra (arXiv:2602.15669)
• Provider research: Persona Vectors, Assistant Axis, Persona Selection Model
• DPRF: Dynamic Persona Refinement Framework (arXiv:2510.14205)
• JPAF: Jungian Personality Adaptation Framework (arXiv:2601.10025)
• Martingale Score: Bayesian Rationality in LLM Reasoning (arXiv:2512.02914)
• Social Sycophancy / ELEPHANT (arXiv:2505.13995)
• Agent Drift / ASI (arXiv:2601.04170)
• PERSIST: Persistent Instability in LLM Personality (arXiv:2508.04826)
• VIGIL: Reflective Runtime for Self-Healing Agents (arXiv:2512.07094)
• Memoria: Scalable Agentic Memory (arXiv:2512.12686)
• TiMem: Temporal-Hierarchical Memory Consolidation (arXiv:2601.02845)
• AgeMem: Unified Memory Management (arXiv:2601.01885)
• SteeM: Controllable Memory Usage (arXiv:2601.05107)
• PATS: Personality-Aware Teaching Strategies (arXiv:2601.08402)
• Provider research: Emergent Misalignment Persona Features (2025)
• Provider research: Constitution update (Jan 2026)
• RULERS: Locked Rubrics and Evidence-Anchored Scoring (arXiv:2601.08654)
• Hindsight: Building Agent Memory that Retains, Recalls, and Reflects (arXiv:2512.12818)
• AgentTrace: Structured Logging for Agent Observability (arXiv:2602.10133)
• CyclicReflex: Cyclical Reflection Token Scheduling (arXiv:2506.11077)
• PreFlect: Prospective Reflection for Agents (arXiv:2602.07187)
• AMDM: Adaptive Monitoring of Agentic AI (arXiv:2509.00115)
• Split Personality Training for LLM Auditing (arXiv:2602.05532)
• Selective Agreement, Not Sycophancy (EPJ Data Science 2025)
• PersonaAgent: Test-Time Personalization (OpenReview 2025)
• SocraticLM: Socratic Personalized Teaching (2025)
• Persona Collapse Taxonomy (HuggingFace 2025)
• SycEval: Evaluating LLM Sycophancy (AAAI/ACM AIES 2025)
• FadeMem: Biologically-Inspired Forgetting (arXiv:2601.18642)
• Belief Decay as Core Mechanism (Artificial Brain Labs 2025)