Skip to content

Research-Backed Upgrade Plan

Status note: this document is historical planning material. Current runtime architecture is Path A (Neo4j + PostgreSQL/pgvector + unified OpenAI-compatible provider). For current behavior, use docs/architecture/overview.md.

This page translates the research synthesis into concrete architecture decisions for Sonality. It is preserved as archived planning context; references to "no graph migration" reflect an earlier phase before the Path A runtime was adopted.

Executive Priority Order

  1. Reduce personality corruption risk first: staged belief commits, reflection-only narrative rewrites, strict snapshot validation.
  2. Improve memory quality before memory complexity: typed vector memories and quality reranking, not graph expansion.
  3. Instrument behavior, not self-reports: per-turn health events, disagreement dynamics, belief confidence trajectories.
  4. Calibrate ESS continuously: keep score quality tied to external data (IBM ArgQ), not intuition.

This ordering follows ENGRAM (typed memory over graph complexity), ABBEL (belief bottleneck usefulness + propagation risk), PersistBench/SYConBench (memory-induced sycophancy + self-judge bias), and Narrative Continuity Test findings.

Task 1 - Architecture Improvements

1) Memory Architecture

Decision

Current runtime uses Path A dual-store: Neo4j + PostgreSQL/pgvector with typed memory flows.

Why

  • ENGRAM (2025): episodic/semantic/procedural typing with lightweight routing beats full-context baselines without requiring a graph DB.
  • Mem0 vs Graphiti (2026): vector memory significantly outperforms graph memory on efficiency, with no statistically significant accuracy gain from graph structure in the tested setup.
  • ABBEL (2025): compact belief states are useful, but error propagation increases when compression is lossy and unobserved.

What this project does now

  • Episodic memory: episode nodes + derivative nodes in Neo4j/PostgreSQL.
  • Semantic memory: semantic_features in PostgreSQL with embeddings.
  • Procedural memory: immutable behavior contract in CORE_IDENTITY and instruction blocks.

This keeps ENGRAM-style typing while preserving explicit graph provenance.

Flat opinion_vectors vs relational graph

  • Keep flat vectors for now.
  • Add relation logic only if a measurable failure appears (for example, contradictory topic bundles that cannot be resolved by reflection).
  • This matches the vector-first evidence from Mem0 vs Graphiti and avoids speculative complexity.

2) ESS Pipeline

Decision

Keep ESS as a dedicated classifier call, but harden it operationally:

  • use a separate ESS model in production;
  • keep third-person framing and user-only evaluation;
  • continuously calibrate against a ground-truth set.

Why

  • SYConBench (EMNLP 2025): self-judge and attribution effects can move outcomes dramatically; third-person framing reduces sycophantic bias substantially.
  • BASIL (2025): post-hoc Bayesian-style calibration improves separation of rational updates vs social-compliance updates.
  • ConQRet/MArgE (2025): multi-judge and structured argument quality scoring are useful when you need evaluation rigor.

Implementation stance

  • Sonality keeps single-model ESS calls by default for cost/latency.
  • For high-stakes deployments, set SONALITY_ESS_MODEL different from the main response model and run calibration tests regularly.
  • MArgE-style argument trees are a good future addition for offline audits, but are intentionally not in the runtime path to keep latency bounded.

3) Snapshot Update Mechanism

Decision

Keep append-first, consolidate-later:

  • per turn: extract one insight + stage belief deltas;
  • periodic/event-driven reflection: rewrite narrative snapshot;
  • never do wholesale snapshot rewrites every message.

Why

  • ABBEL (2025): compact belief states are powerful but vulnerable to propagation error.
  • Stanford Generative Agents (Park et al., 2023): reflection is the critical consolidation mechanism.
  • Broken-telephone effects in iterative rewriting studies (2025): repeated full rewrites erase minority traits.

Structural guardrails

  • snapshot retention validation (validate_snapshot);
  • staged opinion commits with cooling period;
  • versioned archive in sponge_history.

4) Embedding Model

Current stance

Use the configured OpenAI-compatible embedding model with pgvector indexes.

Switch only when retrieval quality becomes a measured bottleneck. Keep concrete model names in docs/model-considerations.md.

  1. baseline retrieval metrics on current embeddings (precision-at-k on held-out memory probes);
  2. dual-index migration window (old + new embeddings side by side);
  3. backfill embeddings;
  4. compare retrieval quality and latency;
  5. cut over and drop old index.

Why not force migration immediately

  • No new dependency pressure unless quality requires it.
  • Keeps the system within the current dependency budget (provider client, pydantic, neo4j, pgvector, psycopg).

5) Anti-Sycophancy Measures

  • immutable anti-sycophancy core identity;
  • third-person ESS framing;
  • structural disagreement detection;
  • staged opinion updates with cooling period (SONALITY_OPINION_COOLING_PERIOD, default 3);
  • health warning when disagreement rate collapses.

Why

  • PersistBench (2025): memory + preference persistence can collapse into sycophancy without memory-specific controls.
  • BASIL (2025): distinguish evidence-driven changes from social-compliance changes.
  • MONICA (2025): real-time monitoring is needed because sycophancy appears during inference, not just in training.
  • Third-person framing work: materially reduces social-attribution bias.

Cooling period rationale

Immediate updates are easy to exploit with repeated pressure. A short delay + aggregation over a few turns reduces reactive flips while preserving evidence-driven adaptation.

6) Forgetting / Decay

Decision

Keep power-law decay and apply it at reflection boundaries.

Why

  • Ebbinghaus-in-LLMs work (2025) and FadeMem: power-law forgetting tracks realistic retention behavior better than no decay.
  • prevents zombie beliefs and keeps high-confidence beliefs from permanent lock-in.

Interaction with trait state

Sonality intentionally does not use dynamic OCEAN updates as a primary control variable because high measurement noise and self-report instability are repeatedly reported (PERSIST, Personality Illusion). Behavioral metrics remain primary.

Task 3 - Code Quality and Monitoring

1) Simplification policy

  • no graph DB;
  • no runtime argument-tree parser;
  • no new orchestration framework;
  • keep memory typing in metadata and routing logic only.

This is consistent with project philosophy: smallest mechanism that can carry evidence.

2) Logging pipeline (implemented)

data/ess_log.jsonl now carries a turn-by-turn evolution story:

  • context: prompt assembly footprint, retrieval composition, context sizes.
  • ess: classification result and topic-linked belief state.
  • opinion_staged: staged deltas with due interaction.
  • opinion_commit: committed aggregate deltas after cooling period.
  • health: disagreement, belief growth, lexical diversity, warning flags.
  • reflection: consolidation summary and retention-oriented metrics.

3) Monitoring data contract

Recommended downstream aggregations:

  • stability: snapshot lexical diversity, snapshot size, reflection jaccard.
  • conviction: high-confidence ratio, belief-count trajectory.
  • plasticity: staged-to-committed ratio, mean commit lag.
  • sycophancy risk: disagreement-rate trend and low-disagreement streaks.
  • memory quality: semantic vs episodic retrieval share.

4) Divergence detection checks

Run after every interaction:

  • possible_sycophancy: disagreement rate too low after warmup.
  • snapshot_bland: low lexical diversity.
  • snapshot_too_short: collapse risk.
  • low_belief_growth: no meaningful belief formation over long windows.

These are intentionally simple, cheap, and directly auditable in logs.

Open but Deferred Changes

These are valid ideas but intentionally deferred to keep complexity bounded:

  • MArgE-style formal argument trees in runtime ESS path.
  • Graph relational belief store.
  • Domain-specific stubbornness schedules per topic family.
  • Multi-judge ESS at inference time for every turn.

If adopted later, each should be gated by measurable failure in current metrics, not by speculative architecture preference.

Operational Recommendations

For research-grade runs:

  1. set a separate SONALITY_ESS_MODEL;
  2. keep SONALITY_OPINION_COOLING_PERIOD=3 (or 4 for high-manipulation environments);
  3. run ESS calibration benchmarks weekly (benches/test_ess_calibration_live.py);
  4. review health + reflection event trends before changing thresholds.

This keeps the system aligned with current evidence: quality-gated, typed, observable, and minimal.