Adaptive Agentic RAG

• Adaptive agentic RAG is a retrieval-augmented framework that explicitly organizes memory access into conditional routing, re-ranking, and synthesis stages.
• It integrates semantic and episodic memory via operations like TRIAGE, DECAY, and CONSOLIDATE to govern persistent memory and support personalized QA.
• The system balances retrieval quality and classification-based re-ranking to enhance factual grounding, mitigate hallucinations, and sustain conversational context.

Searching arXiv for papers on adaptive and agentic RAG. “Adaptive agentic RAG” (Editor’s term) can be understood as a retrieval-augmented generation regime in which retrieval, memory selection, and answer construction are organized as an explicitly controlled, multi-stage process rather than as a single retrieval call followed by generation. In the cited literature, Retrieval-Augmented Generation is described as the dominant pattern for giving LLMs persistent memory, but several adjacent lines of work decompose that pattern into routing, retrieval, synthesis, consolidation, and audit operations. The PerLTQA framework factorizes personalized question answering into Memory Classification, Memory Retrieval, and Memory Synthesis, while “Memory as Metabolism” describes a time-structured memory system with TRIAGE, DECAY, CONTEXTUALIZE, CONSOLIDATE, and AUDIT. Taken together, these works suggest a technically precise view of adaptive agentic RAG as a governed memory-access pipeline in which retrieval is conditioned, re-ranked, consolidated over time, and evaluated for groundedness rather than fluency alone (Du et al., 2024, Miteski, 13 Apr 2026).

1. Conceptual scope and lineage

Adaptive agentic RAG is best situated between two design patterns present in the cited work. The first is retrieval-centric memory access, in which an LLM receives external memories at inference time and synthesizes a response from them. The second is governed persistent memory, in which the memory store itself has lifecycle rules, update schedules, and structural protections. The former is explicit in PerLTQA’s three-stage pipeline; the latter is explicit in the companion knowledge-system design of “Memory as Metabolism” (Du et al., 2024, Miteski, 13 Apr 2026).

This combination matters because the cited work does not treat “memory” as a homogeneous substrate. PerLTQA separates semantic memory from episodic memory, with semantic memory including world knowledge, profiles, and social relationships, and episodic memory including events and dialogues. “Memory as Metabolism” further distinguishes raw buffer entries, active wiki entries, cold memory objects, audit records, and minority branches. A plausible implication is that adaptive agentic RAG should not be defined by the mere presence of a retriever; it should instead be defined by explicit control over which memory type is queried, how retrieved items are fused, when external sources are consolidated, and under what conditions previously dominant interpretations may be revised (Du et al., 2024, Miteski, 13 Apr 2026).

A concise way to summarize the component lineage is as follows.

Component	Mechanism in the cited literature	Source
Routing	Memory Classification	(Du et al., 2024)
Retrieval	cross-type retrieval plus re-ranking	(Du et al., 2024)
Generation	Memory Synthesis via $r' = LLM(z,q,m)$	(Du et al., 2024)
Persistence	wiki-style active and cold memory	(Miteski, 13 Apr 2026)
Governance	TRIAGE, DECAY, CONTEXTUALIZE, CONSOLIDATE, AUDIT	(Miteski, 13 Apr 2026)

The resulting architecture is “adaptive” because retrieval is conditioned and re-ranked rather than fixed, and “agentic” because the system performs explicit operations over memory across time, including consolidation, quarantine, promotion, decay, and audit.

2. Adaptive retrieval as conditional routing and late fusion

The clearest formal account of adaptive retrieval in the cited material appears in PerLTQA. The memory database is defined as

$$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$

and the QA dataset as

$$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$

The pipeline begins with memory classification,

$\pi = MC(q),$

then performs memory retrieval,

$m, s = R(q, M, k),$

and finally applies a re-ranking rule

$$s'_{i} = \alpha \cdot P(\pi|m_i) + \beta \cdot \mathrm{sigmoid}(s_{i}),$$

with $\alpha=\beta=0.5$, before memory synthesis by

$r' = LLM(z, q, m).$

PerLTQA retrieves $k$ memories from each memory category first, collects $2k$ candidates, and only then re-ranks them using both classification probabilities and retrieval score. The appendix states that this design is intended to avoid over-reliance on possibly wrong classification decisions by always retrieving across memory types and then softly re-ranking (Du et al., 2024).

This is a strong operational model for the “adaptive” aspect of adaptive agentic RAG. The adaptation does not depend on a monolithic controller deciding a single memory type in advance. Instead, it arises from soft coupling between a routing model and a retriever. The system first preserves candidate diversity across memory pools and then biases the final context through weighted scoring. The paper explicitly characterizes the resulting integration mechanism as “late fusion through retrieval + prompt conditioning,” not explicit symbolic composition. For adaptive RAG, this implies that robustness may depend less on perfect routing than on a retrieval policy that remains cross-type until late in the pipeline (Du et al., 2024).

The cited results reinforce this interpretation. Removing classification while keeping retrieval causes only small drops, whereas removing external memory causes MAP to fall to roughly $$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$0–$$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$1 and correctness to roughly $$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$2–$$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$3, even though coherence remains high at about $$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$4–$$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$5. The paper therefore concludes that retrieval is the dominant driver, while classification-based re-ranking gives a modest additional gain. In adaptive agentic terms, routing improves selection pressure, but retrieval quality remains the principal bottleneck (Du et al., 2024).

3. Agentic memory governance beyond one-shot retrieval

The most developed account of the “agentic” dimension is the companion-memory governance model of “Memory as Metabolism.” That work explicitly contrasts standard RAG with an LLM wiki pattern that compiles knowledge into an interlinked persistent artifact. In its view, RAG treats each query as a fresh retrieval problem, whereas the wiki pattern builds and governs a structured store whose entries have lifecycle states, dependency relations, gravity, vitality, and audit history. The system is organized around five operations—TRIAGE, DECAY, CONTEXTUALIZE, CONSOLIDATE, and AUDIT—supported by memory gravity and minority-hypothesis retention (Miteski, 13 Apr 2026).

This gives adaptive agentic RAG a temporally extended control logic. TRIAGE is intentionally shallow: it rejects obvious garbage, deduplicates against the recent buffer, checks structural validity, assigns an ingestion timestamp, and assigns a stable content-hash ID. It must not perform semantic contradiction resolution, must not read the active wiki during ingestion, and must not write directly to the active wiki. CONSOLIDATE then performs the deep coherence work on a schedule—nightly, weekly, or event-driven—through four phases: buffer-internal scoring, wiki scoring, classification and routing on a fuzzy coherence gradient, and minority-pressure promotion. AUDIT temporarily suspends top-gravity entries and observes whether query performance degrades, remains unchanged, or improves; it then restores, reduces gravity, or archives accordingly. The design principle is “mirror on operational dimensions, compensate on epistemic failure modes” (Miteski, 13 Apr 2026).

The retention policy is formalized by memory gravity and vitality. The active wiki is $$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$6, and the base gravity of entry $$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$7 is $\pi = MC(q),$5 where $$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$8 is a centrality measure and $$M = \left\{ (S_i(l_1), E_i(l_2)) \mid i = 1, 2, \ldots, p \right\},$$9 is downstream fragmentation cost. The effective gravity is $\pi = MC(q),$6 and the vitality function is $\pi = MC(q),$7 DECAY must apply this vitality formula, must not decay entries whose base gravity remains above the gravity-protection floor, and must compress decay-eligible entries rather than deleting them (Miteski, 13 Apr 2026).

A related but differently formalized control motif appears in MAPO, which stores promising trajectories in a memory buffer and decomposes expected return into expectations inside and outside the buffer:

$$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$0

MAPO supplements this with systematic exploration and memory weight clipping. A plausible implication is that adaptive agentic RAG should treat retrieved evidence as a managed buffer rather than an ephemeral context fragment: once high-value memory items are discovered, they should remain first-class contributors to future decisions instead of depending on accidental rediscovery by the current policy (Liang et al., 2018).

4. Generation as synthesis over retrieved and governed memory

In PerLTQA, memory synthesis is explicitly described as the ultimate goal. The generation stage is prompt-based rather than architecturally elaborate: the LLM receives a prompt template $$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$1, the question $$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$2, and the retrieved memories $$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$3, and produces

$$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$4

The retrieved bundle may include both semantic and episodic items, so synthesis is a late-fusion composition over re-ranked memory snippets. The paper repeatedly stresses that retrieval quality is crucial because the LLM can synthesize well when provided with accurate memories (Du et al., 2024).

This synthesis view is consistent with the governed-memory perspective of “Memory as Metabolism.” CONTEXTUALIZE does not merely compress an external source once; it compresses external sources to the user’s current working-context depth, creates a cold memory object, and preserves a mandatory link to the original source. CONSOLIDATE then decides how new material interacts with the active wiki. A plausible implication is that generation in adaptive agentic RAG should be understood as synthesis over a memory substrate that already encodes editorial decisions about compression depth, source recoverability, contradiction handling, and structural importance, rather than as unconstrained decoding over raw retrieved text (Miteski, 13 Apr 2026).

The guess-and-check pattern described in CNnotator provides another relevant synthesis motif. There, the LLM produces a candidate CN annotation, the system injects it into code, invokes CN’s test-synthesis backend, tests the contract on 100 automatically generated valid inputs, and enters an iterative repair loop if the annotation fails; for the evaluation, refinement is allowed up to $$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$5 times, with special handling for syntax errors. Although this work addresses memory-safety annotation rather than RAG, it offers a concrete model of agentic synthesis as generation conditioned on external context plus downstream checking and repair. This suggests that adaptive agentic RAG need not terminate at first-pass answer generation; it may instead incorporate post-generation verification or repair against retrieved evidence or executable constraints (Byrnes et al., 20 Jun 2026).

5. Evaluation, failure modes, and common misconceptions

The cited work gives a particularly sharp account of how memory-grounded generation should be evaluated. PerLTQA uses correctness, coherence, and MAP of memory anchors, with

$$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$6

This metric is important because it measures whether the synthesized answer actually contains the key memory fragments rather than merely sounding plausible. The principal empirical lesson is that coherence remains high even when answers are not memory-correct. In the setting without any external memory, MAP falls to roughly $$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$7–$$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$8 and correctness to roughly $$T = \{ t_j \}_{j=1}^N,\quad t_j = (q_j, r_j, m_j, a_j).$$9–$\pi = MC(q),$0, but coherence stays high. The paper therefore concludes that the central challenge is accurate memory grounding, not surface generation (Du et al., 2024).

This directly addresses a common misconception: fluent generation is not evidence of successful retrieval augmentation. PerLTQA’s case study illustrates the point with “What is Wang Wei’s occupation?” Without retrieval, the answer hallucinates “teacher”; retrieval without classification yields “actor” from misleading dialogue retrieval; retrieval with classification yields the correct answer “cameraman.” The paper’s interpretation is that synthesis alone is not enough, and retrieval can still mislead synthesis if the wrong memory type dominates (Du et al., 2024).

A second misconception is that adaptive memory systems should integrate contradictory information immediately. “Memory as Metabolism” rejects that approach. TRIAGE must not perform semantic contradiction resolution or direct wiki updates; contradictory evidence should not overwrite central memory in real time, but neither should it be discarded. Instead, contradictory entries accumulate in buffer, quarantine, and minority branches. The system’s sharpest prediction is that accumulated contradictory evidence should have a structural path to updating a centrality-protected dominant interpretation through multi-cycle buffer pressure accumulation, a failure mode the paper states no existing benchmark captures (Miteski, 13 Apr 2026).

A third misconception is that larger memory access alone solves the problem. The cited literature points instead to a governance problem. Retrieval quality, cross-type selection, delayed consolidation, and audit sensitivity all matter. “Memory as Metabolism” is explicit that the safety story is partial: the framework can resist entrenchment, can amplify genuine minority evidence, and can preserve structural continuity, but does not solve reinforcement of bad beliefs, does not guarantee truth, and depends critically on AUDIT sensitivity (Miteski, 13 Apr 2026).

6. Applications, system classes, and open design tensions

The most immediate application class in the cited work is personalized QA. PerLTQA introduces a benchmark of 8,593 questions for 30 characters over semantic and episodic memory, and shows that retrieval-conditioned synthesis substantially outperforms parametric-only answering. Under the best setting with both classification-informed re-ranking and retrieval, $\pi = MC(q),$1 reaches MAP $\pi = MC(q),$2, correctness $\pi = MC(q),$3, and coherence $\pi = MC(q),$4. The semantic-only and episodic-only ablations show that using only one memory type degrades synthesis, supporting the broader claim that explicit cross-memory integration is beneficial (Du et al., 2024).

A second application class is the single-user companion wiki. “Memory as Metabolism” is designed for personal knowledge management, long-lived research or secretarial assistance, companion-style assistants, and AI-assisted software development. Its intended system is a persistent personal knowledge substrate that compiles sources over time, preserves continuity across sessions, and supports personal reasoning and task execution. The paper is especially concerned with entrenchment under user-coupled drift and proposes structural correction channels through scheduled consolidation, minority-hypothesis retention, source-link preservation, no-hard-delete rules, and audit-by-suspension (Miteski, 13 Apr 2026).

These applications reveal the main open tensions in adaptive agentic RAG. One tension is between continuity and revisability: memory gravity protects load-bearing structure, but excessive protection risks the “Absolute Incumbency Trap.” Another is between shallow ingestion and deep integration: premature coherence filtering can create self-sealing memory, but unconstrained retention increases noise. A third is between routing precision and retrieval robustness: PerLTQA’s soft re-ranking is explicitly designed to reduce dependence on classifier precision by retrieving across memory types first. A fourth is between generation and checking: CNnotator’s iterative repair loop suggests that one-shot synthesis is often insufficient when outputs must satisfy external constraints (Du et al., 2024, Miteski, 13 Apr 2026, Byrnes et al., 20 Jun 2026).

In this sense, adaptive agentic RAG is not merely retrieval plus generation. It is a system class in which external memory is typed, routed, re-ranked, synthesized, consolidated, and audited under explicit procedural rules. The cited literature does not present a single unified architecture under that name, but it does provide the constituent mechanisms: conditional retrieval and late fusion from PerLTQA, governed persistent memory from the companion-wiki design, managed buffer logic from MAPO, and guess-and-check synthesis from CNnotator. The resulting technical picture is of retrieval augmentation as a controlled memory metabolism rather than a single retrieval primitive (Liang et al., 2018, Du et al., 2024, Miteski, 13 Apr 2026, Byrnes et al., 20 Jun 2026).