SCH: Stratified Context Hunting Framework

• Stratified Context Hunting (SCH) is a formal framework that addresses context-length limitations using hierarchical memory with three operators: extraction, coarsening, and traversal.
• The framework employs quantitative information-theoretic metrics to guide memory hierarchy design and optimize retrieval performance, ensuring minimal information loss.
• SCH subsumes multiple existing architectures by providing prescriptive rules and bounds, such as Fano-derived partition limits, for efficient context selection under resource constraints.

Stratified Context Hunting (SCH) is a formal framework addressing the context-length limitation in long-context and agentic systems through hierarchical memory. SCH decomposes memory organization and retrieval into three key operators—extraction (α), coarsening (C = (π, ρ)), and traversal (τ)—and introduces quantitative information-theoretic principles to guide the design and analysis of memory hierarchies and retrieval strategies. This formalism subsumes a wide spectrum of existing architectures, providing a unified vocabulary and prescriptive theory for optimizing context selection under resource constraints (Talebirad et al., 23 Mar 2026).

1. Core Components: Extraction, Coarsening, and Traversal

Extraction Operator (α)

The extraction operator α transforms raw or semi-structured data $D \in \mathcal{D}$ (e.g., documents, conversation transcripts, execution traces) into a unit graph $G_0 = (U_0, E_0)$. The atomic information units $U_0$ live in a product feature space $$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$, with one canonical factor $\mathcal{F}_c = \Sigma^*$ holding textual content $\varphi_c(u)$. Edge relations $E_0$ may encode syntactic or semantic relations. The extraction function is defined as

$$\alpha: \mathcal{D} \rightarrow \mathcal{G}, \qquad \alpha(D) = G_0 = (U_0, E_0)$$

where chunking, annotation, or segmentation procedures are encoded within α.

Coarsening Operator ($C = (\pi, \rho)$)

Coarsening reduces the granularity of $G = (U, E)$ to a smaller graph $G_0 = (U_0, E_0)$0 via:

• Partitioning function $G_0 = (U_0, E_0)$1 surjectively groups units, inducing partitions $G_0 = (U_0, E_0)$2.
• Representative function $G_0 = (U_0, E_0)$3 assigns a summary (representative) to each group in the same feature space. Coarsening decreases the number of information units, thus controlling memory footprint. $G_0 = (U_0, E_0)$4 may implement referential labels, abstractive summaries, or embedding centroids.

Traversal Operator (τ)

Traversal selects a subset $G_0 = (U_0, E_0)$5 of atomic units as context, taking as input the hierarchy $G_0 = (U_0, E_0)$6, a query $G_0 = (U_0, E_0)$7, and a token budget $G_0 = (U_0, E_0)$8:

$G_0 = (U_0, E_0)$9

with the constraint $U_0$0. Practical traversal strategies include top-down refinement (beam search), collapsed search (ranking across all layers with expansion), multi-view parallel retrieval, and reasoning-based navigation.

2. Self-Sufficiency of Representatives and Coarsening–Traversal Coupling

The quality of a representative $U_0$1 for a group $U_0$2 is formalized by its self-sufficiency:

$U_0$3

$U_0$4 quantifies the fraction of Shannon information preserved by the representative. High self-sufficiency ($U_0$5) enables retrieval of almost all group information from $U_0$6, whereas low self-sufficiency ($U_0$7) indicates primarily referential summaries.

A query-conditioned variant,

$U_0$8

captures the proportion of query-relevant information retained.

The coarsening–traversal (C–T) coupling is a rate–distortion relationship: if all group representatives satisfy $U_0$9, collapsed search yields at most $$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$0 fraction information loss for any query. For low-SS representatives, only top-down refinement ensures detailed recovery, as collapsed search is insufficient.

Branching factor is additionally bounded via Fano's inequality: the number of partition groups $$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$1 at a coarsened level is limited by the information content of representatives,

$$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$2

to achieve routing error at most $$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$3.

3. SCH Algorithmic Instantiation

A worked SCH algorithm illustrates these principles on a document $$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$4 (e.g., $$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$5 tokens, context budget $$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$6):

1. Extraction (α): Split $$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$7 into $$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$8-token chunks, $$\mathcal{F} = \mathcal{F}_1 \times \cdots \times \mathcal{F}_p$$9, $\mathcal{F}_c = \Sigma^*$0.
2. Hierarchy Construction:
   • $\mathcal{F}_c = \Sigma^*$1: $\mathcal{F}_c = \Sigma^*$2 groups every $\mathcal{F}_c = \Sigma^*$3 chunks (paragraph), $\mathcal{F}_c = \Sigma^*$4 generates a 2-sentence LLM summary per group, producing $\mathcal{F}_c = \Sigma^*$5.
   • $\mathcal{F}_c = \Sigma^*$6: $\mathcal{F}_c = \Sigma^*$7 clusters $\mathcal{F}_c = \Sigma^*$8 into $\mathcal{F}_c = \Sigma^*$9 topics via UMAP+GMM, $\varphi_c(u)$0 generates a single-sentence topic summary per cluster, producing $\varphi_c(u)$1.
3. Traversal (τ) Decisions:
   • If $\varphi_c(u)$2 and $\varphi_c(u)$3 (measured with LLM-proxy $\varphi_c(u)$4), use collapsed search:
     • Pool all units from $\varphi_c(u)$5, rank by query relevance, expand non-leaf nodes, and select atomic units to fill the token budget.
   • If $\varphi_c(u)$6 is low (e.g., only domain labels), perform top-down, stepwise refinement.

4. Comparative Analysis of SCH-Style Systems

The three-operator (α, π, ρ, τ) framework subsumes a broad set of hierarchical memory and retrieval systems. Empirical analysis maps at least eleven architectures, including RAPTOR, xMemory, H-MEM, SimpleMem, GraphRAG, PageIndex, MemoBrain, StackPlanner, AgeMem, InfiAgent, and Om, into this formalism.

Data-memory systems (e.g., RAPTOR, GraphRAG) typically differ in atomic extraction (sentence, entity-pair, episode), grouping schemes (e.g., clustering or structural mapping), representative quality (LLM summary vs. label vs. centroid), and traversal pattern (collapsed vs. refinement). Trace-memory systems (e.g., MemoBrain, StackPlanner) instantiate $\varphi_c(u)$7 over execution steps, coarsen by causal or stack structures, and traverse for execution trace reassembly.

System Domain	Extraction Unit	Coarsening/Representative	Traversal Pattern
Data-memory	sentence, pair, episode	LLM-summary, centroid	collapsed/top-down
Trace-memory	execution step	causal/stack rep	refinement

5. Design Insights and Theoretical Implications

SCH establishes several operational and theoretical principles:

• A unified pipeline: extraction, coarsening, traversal applies across agentic and retrieval-based memory architectures.
• Quantitative metrics (self-sufficiency $\varphi_c(u)$8) link representative quality to retrieval performance.
• Prescriptive design rule (C–T coupling): match high-SS representatives to collapsed search and low-SS representatives to incremental, top-down traversal.
• Fano-derived bounds guide partition granularity based on representative informativeness.
• Affinity or coherence in partitioning ($\varphi_c(u)$9) is necessary for safe pruning during top-down traversal.

6. Limitations, Open Challenges, and Future Directions

Notable limitations of SCH include the assumption of static operators; dynamic, query-conditioned groupings or adaptive refinement of $E_0$0 may violate the separability and Markov assumptions intrinsic to the current formalism. The existing framework does not model the feedback between query-conditioned coarsening and traversal. The theoretical bounds for self-sufficiency and routing rely on idealized Shannon information; practical computation of tight proxies for LLM-based $E_0$1 remains an active area of research (Talebirad et al., 23 Mar 2026).

A plausible implication is that further advances in adaptive information-theoretic modeling and query-conditioned hierarchy construction may extend the prescriptive power of SCH to dynamic, online, and highly agentic language systems.