Agent KB: Architectures and Retrieval

• Agent KB is a framework of structured knowledge bases that dynamically retrieve and integrate agent experiences to enhance tool orchestration and decision-making.
• It employs diverse architectures, including flat vector stores and graph-based schemas, and uses hybrid lexical and semantic indexing for precise retrieval.
• The system supports privacy-aware orchestration and cross-domain generalization, enabling continual updates that improve collective agent intelligence.

Agent KB denotes a class of knowledge-base architectures, retrieval-augmented protocols, and memory infrastructures that equip agentic systems—ranging from tool-using LLM agents to networked multi-agent orchestration frameworks—with structured, dynamically queriable representations of agent experience, tool capabilities, and contextual knowledge. The principal contributions of Agent KB approaches are (1) enabling more robust and generalizable reasoning by grounding agent decisions in externalized, cross-domain knowledge, (2) facilitating cross-agent and cross-architecture knowledge transfer, and (3) supporting fine-grained, privacy-aware orchestration, access control, and planning. These functions are realized through a spectrum of formal schemas (flat vector-indexed stores, experience graphs, bipartite agent–tool knowledge graphs), retrieval pipelines (hybrid lexical/semantic scoring with safety gates), and integration protocols that are collectively establishing the foundational infrastructure for practical collective intelligence in heterogeneous agentic systems (Tang et al., 8 Jul 2025, Nizar et al., 22 Nov 2025, Trombino et al., 23 Sep 2025, Zhou et al., 15 Nov 2025).

1. Architectures and Schemas of Agent KBs

Agent KB implementations range from simple flat vector stores to bipartite graphs and multimodal, hierarchical repositories. For instance, universal memory systems such as "Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving" (Tang et al., 8 Jul 2025) aggregate agent trajectories into structured “experience units” $E = \langle\pi, \gamma, S, C\rangle$ where

• $\pi$ is a dense task (plan) embedding,
• $\gamma$ encodes structured goal predicates,
• $S$ is a log of action–reasoning pairs,
• $C$ is framework-specific compatibility metadata.

These units are stored in hybrid indices (BM25 + dense k-NN vector search) and exposed via lightweight APIs.

In contrast, graph-based approaches such as "Agent-as-a-Graph" (Nizar et al., 22 Nov 2025) model agents and their toolkits as a bipartite directed knowledge graph $G=(V,E), V=A\sqcup T, E\subset T\times A$, with agent and tool nodes richly annotated and linked via explicit “owns” edges. Query-time retrieval thus fully captures both fine-grained tool capabilities and agent-level contextual constraints.

Other domains, including authorization for intent-based management (Abdelrazek et al., 22 Oct 2025), employ KBs as semantic knowledge graphs (RDF) whose access rights are governed by dynamic intersectional policies over agent roles, contextual/functional attributes, and predicate-level permissions.

2. Aggregation and Indexing Mechanisms

Experience aggregation in Agent KBs entails continuous ingestion of agent trajectories or human demonstrations, accompanied by abstraction into reusable knowledge units. Indexing typically proceeds on dual axes:

• Lexical Indexing: BM25/textual inverted indices enable sparse matching.
• Semantic/Vector Indexing: Task descriptions, plans, tool functions, and experience logs are embedded via encoders such as all-MiniLM-L6-v2 or OpenAI ada-002. Cosine similarity is used for dense semantic search.

Hybrid fusion, e.g., $$\sigma_i^\mathrm{hyb} = \alpha \tilde{\sigma}_i^\text{text} + (1-\alpha)\tilde{\sigma}_i^\text{sem}$$, is employed for candidate ranking; type-specific weighting yields further gains in retrieval accuracy (Tang et al., 8 Jul 2025, Nizar et al., 22 Nov 2025).

Graph-based KBs support traversal-based expansion (e.g., tool$\to$agent rollup), admitting further structural filtering and context-sensitive retrieval (multi-hop expansions, metadata constraints).

3. Retrieval, Planning, and Diagnostic Feedback

Agent KB retrieval protocols universally adopt a two-stage hybrid pipeline (Tang et al., 8 Jul 2025, Zhou et al., 15 Nov 2025, Nizar et al., 22 Nov 2025):

• Planning Seed Stage: Task descriptions are embedded and matched against KB trajectories (or agent/tool nodes). From top-k retrieved experiences, the agent adapts planning patterns using available tools/context via provided metadata.
• Feedback/Repair Stage: Post-execution error summaries or failure states are matched in the KB to identify corrective patterns—which are gated by plan similarity (the “disagreement gate” $$G(\rho, \rho')=\mathbb{1}[\cos(\phi(\rho),\phi(\rho'))\geq\beta]$$, typically $\beta=0.8$)—to avoid catastrophic interference in heterogeneous settings.

Reinforcement is achieved through runtime updates to the KB as new experiences are encountered and validated, supporting continual integration of fresh, framework-agnostic problem-solving strategies.

4. Privacy, Authorization, and Orchestration

Modern Agent KB protocols increasingly emphasize privacy-preserving orchestration and access control, particularly in multi-agent systems with sensitive internal knowledge. "Knowledge Base-Aware Orchestration" (Trombino et al., 23 Sep 2025) introduces a privacy-respecting relevance protocol: each agent, upon receiving a task probe, computes a local relevance score $s_i(T)$ over its internal KB, returning only a minimal {OK, KO, Partial} signal. The orchestrator aggregates these signals and caches task–agent matches using a semantic cache (vector-keyed store with time-to-live and invalidation via KB change broadcasts), achieving large empirical gains in routing precision while preventing KB data exfiltration.

Agent KBs in network management are further secured by contextual-attribute and functionally-aware access control. The authorization model $\mathcal{M}=(A,R,C,F,KG,P,\phi)$ realizes least-privilege enforcement via intersectional policy composition, default-deny semantics, and per-request reevaluation in dynamic contexts (Abdelrazek et al., 22 Oct 2025).

5. Role in Cross-Domain Generalization and Collective Intelligence

Agent KBs are the enabling substrate for cross-domain knowledge transfer and collective agent intelligence. Notably, the universal memory layer proposed in (Tang et al., 8 Jul 2025) demonstrates interoperability across smolagents, OWL, SWE-Agent, and OpenHands, allowing experience harvested in one system to seed inference and repair in others—without retraining or architectural changes. Empirical evaluation shows substantial improvements across diverse LLM/model families and domains, with pass@3 gains up to +18.7 percentage points on code/QA benchmarks.

The introduction of dynamic, actively retrieved planning seeds and feedback is essential for performance: ablations consistently show 1–7 percentage point regressions when these mechanisms are disabled. Cross-architecture generalization remains strong; however, limitations exist in asymmetric domain transfer and retrieval plateauing at high KB volumes, underscoring the need for richer structuring and learned retrieval in future work.

6. Specialized KBs for Hierarchical and Multimodal Agents

Task success in hierarchical and multimodal agentic settings crucially depends on careful KB stratification and retrieval granularity. In mobile and robotic automation, for example, Mobile-Agent-RAG (Zhou et al., 15 Nov 2025) employs two distinct, retrieval-augmented knowledge bases: Manager-RAG for high-level, human-verified plans, and Operator-RAG for low-level UI or control primitives. This explicit knowledge decoupling yields double-digit improvements in task completion rates and step efficiency for long-horizon, multi-app workflows.

In multimodal conversational agents, late-stage fusion of compact KB embeddings with hierarchical context dramatically improves response quality on knowledge-dependent intents, as empirically demonstrated in (Agarwal et al., 2018).

7. Performance Characteristics and Future Directions

Agent KB architectures consistently outperform description-only or local-memory baselines. For instance, bipartite Agent-as-a-Graph retrieval outperforms standard agent and tool retrievers by 14.9–19.4 percentage points Recall@5 and 14.6–17.7 points nDCG@5 across embedding models (Nizar et al., 22 Nov 2025). KBA orchestration achieves up to +43.5 point gains in routing accuracy vs. minimal baselines (Trombino et al., 23 Sep 2025). Robustness to noisy or inconsistent KBs is addressed by methods such as DKAF, which arbitrates dynamic snapshots via insertion/deletion/completion modules to maintain learning reliability in volatile domains (Saley et al., 2023).

Emergent directions include learning-based hybrid retrieval fusion, dynamic consolidation/pruning in large-scale experience graphs, support for multimodal and structural knowledge (e.g. code ASTs, tool logs), and deeper integration with hierarchical planning and symbolic reasoning frameworks (e.g. DSPy) (Tang et al., 8 Jul 2025). These advancements aim to address scalability, generalizability, and contextual sensitivity, solidifying Agent KBs as central infrastructure for future large-scale, agentic AI systems.