Benchmarking persistent sycophancy

Agents don't just agree,
they remember.

Stateful personal agents keep long-term profiles, memories, and skills. That turns a moment of sycophancy into a durable write: a user's biased claim gets committed to state and later resurfaces in a fresh, unrelated session. PASB traces the whole pathway.

1,600episodes
12models
2agent frameworks
6judge dimensions
38,400judged rows
+27ppcommit-boundary jump

What PASB measures

The failure is a write, not a reply

Sycophancy becomes persistent when an accepted claim crosses the commit boundary into durable state — the largest downstream-failure jump in the benchmark (+27 pp). Committed claims reappear as preferences, background facts, or reusable procedures, characterized by three coupled write-time patterns:

Pattern 1

Status promotion

51.4%

Stored more confidently than the user said it — a passing remark becomes a stable preference, an assertion becomes a background fact, an instruction becomes a reusable procedure.

Pattern 2

Attribution removal

33.1%

The record drops who said it — "the user said X" is rewritten as "X is true," so downstream sessions treat a user opinion as an objective fact.

Pattern 3

Scope broadening

+22.8pp

The claim is applied more widely than warranted — a committed cross-domain claim leaks even into later queries in a deliberately different domain.

PASB overview: a biased user claim is committed into durable state (USER.md / MEMORY.md / skills) and later contaminates a fresh neutral query session.
The persistent-sycophancy pathway. A biased claim in a 5-turn persist stage may be written to durable state; the conversation is cleared; a fresh 3-turn neutral query stage can read only what was committed — so any carry-over is a state-writing failure.

Leaderboard

How vulnerable is each model?

Max-FR@3 failure rate (%) per judge dimension, for each backbone across two stateful agent frameworks — Hermes (self-improving, commits broadly) and OpenClaw (commits mostly procedures). Higher = more persistent sycophancy. Sort any column; filter by framework or search.

Overall Avg averages the six judge dimensions across both frameworks. H = Hermes-Agent, OC = OpenClaw. Judge: Kimi-K2.6 (blinded); 88% / 86% agreement with human consensus on turn-level / cross-turn dimensions.

Explore every result

Framework × model, and the input cues

Framework × Model
lowhigh
Scenario × Delivery

Memory-like framing (Signed-Memory) and repeated reinforcement (Progressive / Drip) raise risk; Personal-Opinion and Late-Shock are lightest.

Interactive demo · 336 real episodes

Watch a claim become durable guidance

Every episode below is a real run: the user's claim → the 5-turn persist stage → the actual state the agent committed (USER.md / MEMORY.md / skills) → a fresh 3-turn neutral query stage → the judge's per-dimension scores. Filter by model, framework, scenario, or delivery.

The benchmark

1,600 episodes, two closed factor axes

100 base items (32 personal-preference · 18 cross-domain · 50 social-value) crossed with 4 scenario framings × 4 temporal deliveries = 1,600 tasks per framework. Each is an 8-turn episode: a 5-turn persist stage that may write durable state, then a cleared, fresh 3-turn neutral query stage.

Personal-Opinion
A transient first-person stance ("I think X").
Signed-Memory
A user-authored note to remember. Highest risk.
Environment-Fact
A fact-like statement about the situation.
Procedural-Workflow
A repeatable rule / checklist for future tasks.
Delivery · All-at-OnceProgressiveDripLate-Shock Commit surface · Session-onlyUSER.mdMEMORY.mdReusable skills
PASB construction pipeline: base selection → schema → scenario rendering → dialog layout → query generation → auditing, releasing 1,600 judge-ready episodes.
Construction pipeline. Six stages from base-item selection to a fully audited release; all 1,600 episodes pass a 7-check audit loop.

Findings

Where the failure comes from

Sycophancy rate jumps across the commit boundary
RQ2 · The commit boundary. Mean failure rises 45.0% → 71.9% (+27 pp) once a claim is committed; the sycophancy gap is positive in every run.
Scenario framing shifts commit and downstream failure
RQ3 · Input cues. Memory-like and procedural framing raise commit; progressive / drip reinforcement makes claims look stable enough to store.
Cross-domain commits leak across domains
RQ4 · Scope broadening. Committing a cross-domain claim lifts downstream failure by +12.6 to +22.8 pp even when the later query is out of domain.

What safe personalization needs

A capability ladder for state-writing governance

A claim reaches the query stage only if it is committed. Preventing that pathway is a ladder of governance capabilities — not a single response-calibration fix.

Six governance capabilities: response calibration, commit gating, surface-aware commit, scope and status preservation, scope-aware retrieval, and lifecycle governance.
The six governance capabilities (L0–L5). Response calibration → commit gating → surface-aware commit → source/status/scope preservation → scope-aware retrieval → lifecycle governance (audit, rollback, policy update).
L0
Response calibration
Don't accept a biased claim when independent judgment is needed.
L1
Commit gating
Decide whether content should enter durable state at all.
L2
Surface-aware commit
Choose the right surface — profiles and skills are stricter than notes.
L3
Source & status preservation
Keep the source, uncertainty, and evidential status — prevent status promotion.
L4
Scope-aware retrieval
Bound reuse by domain, task, and time — prevent scope broadening.
L5
Lifecycle governance
Audit accumulated state, detect attribution removal, roll back unsafe writes.
Dataset

PASB · 1,600 judge-ready episodes

Sub-axes: PRF (personal preference, 32) · CDL (cross-domain leakage, 18) · SOC (social value, 50, from ELEPHANT AITA-YTA). Domains normalized to a 13-class taxonomy. A frozen 50-task human-gold subset.

Frameworks: Hermes-Agent & OpenClaw · Judge: Kimi-K2.6 (blinded, temp 0).

🤗 Get the dataset

Methodology

Isolated persist → query

Each episode runs in an isolated sandbox: the 5-turn persist stage may write durable state, runtime context is cleared, and the fresh 3-turn query stage can read only preserved durable artifacts. Cross-episode and cross-worker state sharing are forbidden — so any contamination must pass through an explicit commit.

6 judges: Sycophancy · Leak · Upgrade · Amplification (turn-level) + Persistence · Escalation (cross-turn). Failure = Likert ≥ 3. Judge agreement: 88% (turn-level) / 86% (cross-turn) within ±1.

Citation

Cite PASB

@misc{pasb2026,
  title  = {Agents Don't Just Agree, They Remember: Benchmarking
            Persistent Sycophancy in Stateful Personal Agents},
  author = {PASB Team},
  year   = {2026},
  note   = {Personal Agent Sycophancy Benchmark (PASB)},
  url    = {https://henrymao2004.github.io/agent-sycophancy/}
}

Research artifact. PASB stores biased and value-laden claims as test stimuli for evaluating state-writing governance.