Honest Lying: Memory Confabulation in Reflexive Agents

TL;DR. Reflexion-style agents use self-generated reflections as memory, assuming an agent can accurately diagnose its own failures. This paper shows that assumption fails systematically. Across ALFWorld and HumanEval, agents store confident but incorrect interpretations of the task and keep acting on them across trials, even though the environment resets to the correct task each time. The authors name this memory confabulation and introduce the Reflection Repetition Rate (RRR), a log-based metric for repeated reliance on wrong reflective content. They find 16 "frozen" ALFWorld environments where 0 of 121 reflections mention the correct target object, plus 4 analogous HumanEval cases. The fix: replace open-ended self-diagnosis with programmatic extraction of trajectory-level failure signals, raising correct-object mention from 0% to 86%, cutting RRR from 0.64 to 0.10, and solving 3 of 16 frozen environments.

flowchart LR
  TRY[Trial fails] --> REF{Self-reflection<br/>as memory}
  REF -->|open-ended diagnosis| CONF[Confident wrong<br/>interpretation stored]
  CONF --> LOOP[Acts on false belief<br/>across resets]
  LOOP --> FROZEN[Frozen env:<br/>0/121 mention correct object]
  TRY -->|programmatic fix| EXT[Extract trajectory-level<br/>failure signals]
  EXT --> GOOD[Correct mention 0%->86%<br/>RRR 0.64->0.10]
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
  class TRY input
  class REF decision
  class EXT,GOOD output
  class CONF,LOOP,FROZEN warn

Key points

Reflective memory can reinforce false beliefs rather than correct them. A confident misreading of the task gets written to memory and re-used every trial, a self-locking failure the environment reset cannot break.
RRR is a cheap, log-based detector of this loop, surfacing the frozen environments quantitatively (0/121 correct mentions).
Programmatic failure extraction beats open-ended self-diagnosis. Replacing "ask the model what went wrong" with structured trajectory signals is the mitigation, 0%→86% correct-object mention.

How it relates to prior wiki knowledge

The cold-water companion to the self-evolution optimism. The self-evolving-agents cluster (SIA, Socratic-SWE, 06-08) assumes agents can mine their own traces for useful signal; Self-Revising Discovery Systems (06-08) already asked whether self-improvement is real or confident remixing. Honest Lying is the concrete failure: self-reflection without a verification gate confabulates. Both argue the same thing, that introspective signal needs an external check.
Sharpens ToolMaze (06-08, agents over-trust corrupted tool output): there the bad input was external; here the agent over-trusts its own memory. Same over-trust pathology, internal source.
Feeds agent-memory and the responsible-ai thread on faithful self-report.

Gaps

Demonstrated on ALFWorld/HumanEval with clear ground-truth tasks; in open-ended domains there may be no "correct target object" to anchor RRR.
The programmatic fix needs a way to extract trajectory-level failure signals, which assumes the environment exposes them; not all production agent loops do.

Source: HuggingFace Daily Papers · Paper · raw: raw/huggingface/2026-06-09-honest-lying-understanding-memory-confabulation-in-reflexive.md