Structured Success Memory Profile

• Structured Success Memory Profile is a formal framework that defines memory organization, abstraction, and retrieval to determine agent performance in complex tasks.
• It specifies a taxonomy of representations, retrieval algorithms, and quantitative success metrics, integrating techniques like iterative retrieval and hierarchical abstraction.
• The framework guides robust design trade-offs and adaptive strategies, with implications for improved multitask performance in language QA, tracking, and evolutionary settings.

A Structured Success Memory Profile (SSMP) is a formal characterization of how the organization, abstraction, and retrieval of memory determines agent performance, adaptation, or evolutionary fitness in complex tasks. SSMPs span domains including long-horizon language-agent tasks, evolutionary game theory, meta-cognitive agent architectures, and recoverable tracking, unifying them via their focus on explicit memory structuring as a determinant of task-specific and environmental success. SSMPs typically specify: (i) a taxonomy of memory representations, (ii) memory abstraction mechanisms, (iii) retrieval and update algorithms, (iv) quantitative success criteria, and (v) task-environment mappings to optimal memory strategies.

1. Taxonomy and Formalization of Structured Memories

Recent research (Zeng et al., 2024, Liang et al., 12 Jan 2026, Huang et al., 2024) formalizes structural memory by encoding domain-relevant experience into explicit, modular representations:

• Chunks: Fixed-length, contiguous text or signal segments used to preserve low-level sequential context. Notation: $\mathcal{C}_q = \{c_1, \ldots, c_J\}$, $|c_j| \leq L$ (Zeng et al., 2024).
• Knowledge Triples: (head, relation, tail) tuples capturing semantic or relational dependencies, generated by extraction models or LLMs (Zeng et al., 2024).
• Atomic Facts: Minimal, self-contained declarative units (often sentences) (Zeng et al., 2024).
• Summaries: High-level, semantically compressed overviews optimized for long-context or narrative retention (Zeng et al., 2024).
• Mixed: Union of above types to balance fine-grained recall with high-level abstraction (Zeng et al., 2024).
• Hierarchical Abstractions: Multi-level representations parameterized by abstraction level $\alpha$, mapping execution trace $\tau$ to memory $m_\tau(\alpha)$. Memory hierarchies $\mathcal{H} = \{H_0, \ldots, H_L\}$ enable selective retrieval at episodic (low $\alpha$) to schematic (high $\alpha$) levels (Liang et al., 12 Jan 2026).
• PN Tree Structures: Tree memory maintaining separate, bounded positive (success) and negative (failure) branches to preserve recency-ordered evidence for robustness and self-recovery (Huang et al., 2024).

This formalization allows memory selection and manipulation to be decoupled from task logic and adaptive over time or evolutionary epochs.

2. Memory Abstraction, Optimization, and Retrieval Algorithms

Memory management is operationalized through both rule-based and learnable methods:

Optimization and Abstraction

• Learnable Abstractions: MCMA (Liang et al., 12 Jan 2026) introduces memory abstraction as a meta-cognitive skill, parameterized by $\alpha \in [0,1]$, selecting representations along the abstraction spectrum. Abstractions are learned via direct preference optimization (DPO), maximizing downstream task success by constructing preference pairs $(m^+, m^-)$ and minimizing

$$L_{DPO}(\theta) = -\log \sigma\left(\beta \left[\log p_\theta(m^+\mid\tau) - \log p_\theta(m^-\mid\tau)\right]\right)$$

so that the memory copilot produces structures that support future agent performance (Liang et al., 12 Jan 2026).

Retrieval Algorithms

• Single-step retrieval: Embedding-based ranking of memory candidates by query similarity for top-$K$ selection (Zeng et al., 2024).
• Reranking: LLM-based refinement of an initial candidate set, selecting top performers via relevance scores (Zeng et al., 2024).
• Iterative retrieval: Query refinement through successive retrieval and LLM-aided reformulation over $N$ steps, with enhanced support for complex, multi-step tasks (Zeng et al., 2024).
• Tree-walking (PN Trees): Branch-based traversal with cosine similarity for robust classification of state as success/present or failure/absent, maintaining relative, not fixed, similarity thresholds (Huang et al., 2024).

3. Quantitative Criteria and Empirical Evaluation

Success is formalized by explicit metrics, depending on domain:

• Language QA and Dialogue (Zeng et al., 2024):
  • Exact Match (EM): Binary hit/miss of answers.
  • F1: Token-level precision/recall harmonic mean.
  • Accuracy (ACC): Strict correctness for reading comprehension.
• Tracking (Huang et al., 2024):
  • Precision, Normalized Precision, AUC, Recovery Rate $SR(T)$ (fraction of lost targets recovered within $T$ frames).
  • Short-term success rate $R_s = N_p / (N_p + N_n)$ over recent memory window.
• Evolutionary Success (Sun et al., 13 Sep 2025):
  • Cooperation rate $C$ as a function of benefit-to-cost ratio $b/c$ and network parameters.
  • Unifying indicator $\mathcal{L}_{Xn}^Y(\delta, \epsilon)$, quantifying leverage from memory length and retrieval policy, with analytic connection to the critical $(b/c)^*_{Xn}^Y$ threshold.

Empirical benchmarks consistently demonstrate that structural memory design, memory abstraction, and retrieval strategy jointly control performance envelopes: e.g., Mixed+Iterative yields F1 $\geq 82\%$ in multi-hop QA (Zeng et al., 2024); MCMA raises ALFWorld success from $55\%$ (no memory) to $79.29\%$ (Liang et al., 12 Jan 2026); RTracker achieves $SR(10 \text{ frames}) \approx 80\%$ vs baselines $< 65\%$ (Huang et al., 2024).

4. Task-to-Strategy Mapping and Phase Diagrams

The SSMP paradigm associates specific memory structures and abstraction/retrieval regimes to task characteristics and environment:

Task/Domain	Structure + Retrieval/Abstraction Policy	Success Metric / Benchmark
Multi-hop QA (HotPotQA)	Mixed + Iterative, Atomic+Iterative (alt)	F1 $\geq$ 82%
Multi-hop QA (2Wiki)	Mixed + Iterative, Triples+Iterative (alt)	F1 $\geq$ 68%
Single-hop QA (NarrativeQA)	Summaries + Iterative	F1 ≈ 31–33
Dialogue Understanding (LoCoMo)	Chunks + Iterative	F1 ≈ 45.1
Reading Comprehension (QuALITY)	Chunks + Reranking or Mixed+Iterative	ACC $\approx$ 79.5%
Long-horizon LLM planning (ALFWorld, etc.)	MCMA: Hierarchical abstraction, DPO-optimized copilot	Success $\uparrow$ 24–28% (absolute)
Social Evolution (network, high $\delta$)	Long-memory $(n\gg1)$, stochastic/deterministic policy	Minimum $b/c^*$ threshold
Noise/High Distribution Shift	Mixed memory, PN trees, or transferable abstraction skill	Success degrades slowest

Phase diagrams in the evolutionary setting partition parameter space $(\tau, \delta, \epsilon)$:

• Low error, high discount ($\delta \approx 1$): use long memory.
• Higher noise/turnover: revert to reactive, memory-1, or even memoryless strategies (Sun et al., 13 Sep 2025).

5. Practical Implications: Design, Robustness, and Trade-Offs

Design guidelines derived from SSMP analyses align structural memory with both computational and task-driven constraints (Zeng et al., 2024):

• Leverage relational representations (Triples, Facts) for reasoning; combine in Mixed for precision/recall trade-off.
• Select Chunks or Summaries for long-context or narratives where high-level abstraction suffices.
• Prefer Iterative retrieval for multi-step reasoning, Reranking for refinement, Single-step for speed-sensitive or simple tasks.
• Choose answer-generation mode: Memory-Only maximizes precision; Memory-Doc (full context) supports more holistic tasks.

Trade-offs:

• Mixed memory incurs storage/retrieval overhead but excels under distributional noise.
• Iterative retrieval increases latency but improves accuracy.
• Summaries economize on storage but risk omitting critical details.
• In MCMA, transferable abstraction skills (copilot parameter $\theta$) enable generalization to OOD tasks even without data-level memory transfer (Liang et al., 12 Jan 2026).

For tracking, PN trees balance recovery and robustness, maintaining a fine-grained, temporally ordered record of successful and failed observations that supports rapid, reliable re-initialization after occlusion or drift (Huang et al., 2024).

6. Extensions and Research Directions

Open directions identified in SSMP research include:

• Dynamic memory consolidation: Pruning, merging, or adapting memory structures over agent lifespan (Zeng et al., 2024).
• Hybrid/adaptive retrieval: Learning to dynamically switch between retrieval routines based on task complexity or resource constraints (Zeng et al., 2024).
• Self-evolving or multi-agent memory: Exploring shared, compositional, or tool-augmented memory modules (Zeng et al., 2024).
• Learnable abstraction transfer: Transferring not only abstracted memories, but also the meta-cognitive skill of abstraction (as in copilot transfer, MCMA) across disjoint tasks or domains (Liang et al., 12 Jan 2026).
• Robust tracking frameworks: Advanced tree-based or graph-based memory management to further improve re-identification and temporal continuity under prolonged or repeated occlusion (Huang et al., 2024).

A plausible implication is that SSMPs provide a unified formalism for predicting, explaining, and optimizing memory-centric agent behavior in both artificial and natural settings, with immediate consequences for the development of resilient, adaptive, and transferable AI systems.