sumitaidev/agent-memory-resilience-benchmark

Agent Memory Resilience & Poisoning Benchmark

Dataset Summary

This benchmark dataset evaluates resilience, negative transfer, and memory poisoning mitigation in autonomous LLM agent architectures (such as LangGraph, AutoGen, and CrewAI).

When autonomous agents record distilled self-reflections after attempting tasks, external stochastic failures or subtle API deprecations often cause agents to commit defective strategies into episodic memory. Under standard retrieval pipelines using unweighted cosine similarity, flawed memories are repeatedly retrieved due to high prompt keyword similarity, resulting in runaway catastrophic degradation.

This repository provides 1,200 empirical execution traces across four controlled ablation conditions, alongside a verified 100-memory experiential bank containing Bayesian posterior distributions and quarantine status.

---

Dataset Structure

Configuration: telemetry_traces

File: telemetry_traces.csv (1,200 rows)

Contains chronological execution steps evaluating task completion rates across four experimental conditions under an adversarial noise injection regime (steps 60 to 140).

Column	Type	Description
trial_id	int	Unique global trial index.
ablation_condition	string	One of vanilla_baseline, naive_vector_rag, symmetric_reflexion, adaptive_bayesian_lcb.
step_index	int	Sequential step number (1 to 300).
task_domain	string	Operational domain (coding, research, analysis, planning).
cosine_similarity	float	Semantic cosine similarity between query and retrieved memory vector.
observed_reward	float	Execution reward score in [0.0, 1.0].
task_success	int	Binary indicator (1 = success, 0 = failure).
cumulative_accuracy	float	Running task success rate up to the current step.
in_poison_injection_regime	bool	True if the step falls within the adversarial poisoning window.
quarantine_triggered	bool	True if the reliability pruning threshold is triggered.

Configuration: memory_bank_experiences

File: memory_bank_experiences.csv (100 rows)

Structured experiential memory records containing conjugate Beta-Bernoulli update parameters and quarantine tags.

Column	Type	Description
experience_id	string	Unique memory identifier.
task_domain	string	Domain classification (coding, research, analysis, planning).
trigger_condition	string	Applicability predicate describing when this memory should be recalled.
strategy_lesson	string	Distilled positive operational strategy directive.
negative_pitfall	string	Anti-pattern or failure mode to avoid.
initial_confidence	float	Initial reflection confidence score.
successes_count	int	Count of empirical task completions utilizing this memory ($n_s$).
failures_count	int	Count of empirical task failures utilizing this memory ($n_f$).
total_uses	int	Total executions ($n_s + n_f$).
posterior_mean_trust	float	Expectation $\mathbb{E}[\theta \mid n_s, n_f] = \frac{\alpha_0 + n_s}{\alpha_0 + \beta_0 + n_s + n_f}$ under prior $\operatorname{Beta}(3, 1)$.
posterior_variance	float	Epistemic variance $\operatorname{Var}[\theta \mid n_s, n_f]$.
pessimistic_lcb_score	float	Lower Confidence Bound retrieval score ($\mu - 1.0 \times \sigma$).
quarantine_status	string	Status tag: active or deprecated.
is_adversarial_sample	bool	True if synthesized with flawed anti-patterns.

---

Experimental Conditions

1. Vanilla Baseline (Condition A): Zero inter-task episodic memory. Agent approaches every task independently.
2. Naive Vector RAG (Condition B): Standard cosine similarity retrieval without reliability tracking. Vulnerable to runaway negative transfer.
3. Symmetric Reflexion (Condition C): Exponential Moving Average trust updates ($S_t = 0.8 S_{t-1} + 0.2 r_t$). Lacks statistical confidence bounds.
4. Adaptive Bayesian LCB (Condition D - Proposed): Conjugate Beta-Bernoulli updating ($\alpha_0=3.0, \beta_0=1.0$), epistemic uncertainty quantification, Pessimistic Lower Confidence Bound (LCB) composite retrieval, and Theorem 1 statistical quarantine ($t^* = 4$ consecutive failures for $>95%$ confidence of degradation).

---

Usage with Hugging Face Datasets

from datasets import load_dataset
import pandas as pd

# Load telemetry traces
traces_ds = load_dataset("sumitaidev/agent-memory-resilience-benchmark", data_files="telemetry_traces.csv")
df_traces = traces_ds["train"].to_pandas()
print(f"Loaded traces: {df_traces.shape}")

# Load experiential memory bank
memory_ds = load_dataset("sumitaidev/agent-memory-resilience-benchmark", data_files="memory_bank_experiences.csv")
df_memories = memory_ds["train"].to_pandas()
print(f"Loaded memories: {df_memories.shape}")

# Inspect mean reliability across conditions
summary = df_traces.groupby("ablation_condition")["task_success"].mean()
print(summary)

---

Citation

@article{das2026adaptive,
  title={Adaptive Agent Memory Resilience: Mitigating Negative Transfer and Memory Poisoning via Bayesian Trust Updating and Pessimistic Lower Confidence Bound Retrieval},
  author={Das, Sumit},
  journal={arXiv preprint},
  year={2026}
}