Toward Knowledge-Centric AI: Discovering, Internalizing, and Using Explicit Knowledge Across Agents and Generations

Yisen Gao, Haoyu Huang, Zhongwei Xie, Yufei Li, Hong Ting Tsang, Boyi Zhang, Qiao Xiao, Tianshi Zheng

Nov 2, 2026arXiv:2611.03157v1

cs.AI

Scorecard· 12/16

6.0/10 impact

Well-argued, timely position paper with concrete invariants and a falsifiable agenda, but no empirical validation and substantial overlap with existing KG/agent-memory/large-knowledge-model framings.

Abstract

Knowledge is how understanding accumulates across people and generations: durable records that outlive their producers and that others can check, combine, teach, and correct. AI agents are now built and run by many independent users and organizations, and each agent is upgraded, replaced, or retired within months. Yet what an agent learns stays trapped in model parameters, a transient context window, or a system-specific memory store that is appended to and retrieved from rather than processed. We argue that the next generation of AI should be knowledge-centric, built around an explicit, continuously evolving knowledge layer independent of any single agent. In this loop, agents discover knowledge from their environments and trajectories; the layer internalizes it by turning fragmented, heterogeneous records and experience into composable, revisable understanding; and agents across systems and generations use the result and return new outcomes for revision. While discovery and use have advanced rapidly with large language model agents, internalization remains the bottleneck. We argue that internalization should become a persistent object of training and maintenance-centered on condition-preserving conceptualization of events and deductive, inductive, and abductive reasoning across noisy, heterogeneous representations (text, tables, graphs, rules, and programs). To power this layer, we call for knowledge foundation models that provide shared neural capacity for knowledge computation, and outline a research agenda showing how an explicit knowledge layer delivers reuse, cost efficiency, auditability, and continuity across agent generations.

AI Impact Assessments

(1 model)

Paper type: Position/perspective paper with a literature-based gap analysis. It contains no original experiments. Its only quantitative figures are cited from prior work.

Core Contribution

The paper argues that AI should be "knowledge-centric": an explicit, agent-independent knowledge layer that agents discover knowledge into, that is internalized, and that agents of other systems and later generations use. Its main conceptual moves are:

A knowledge-unit tuple κ=(content, form, conditions, provenance, status).

A discovery–internalization–use loop.

Isolating "internalization" as the neglected stage, with a bottleneck hypothesis. This stage covers condition-preserving conceptualization, cross-carrier alignment, and deductive, inductive and abductive reasoning.

Five operator invariants (I1–I5): provenance preserved, independent-source support counting, role/type/unit compatibility, conditions and status not silently dropped, and retraction propagation.

A working objective (Eq. 2).

A call for "knowledge foundation models" (KFMs).

Three named gaps and a seven-challenge research agenda.

A proposed critical baseline: same-model agents with raw records, retrieval and code execution, compared against a processed knowledge layer under equal budgets.

Much of the underlying territory is not new. It overlaps with knowledge graphs, agent memory, large knowledge models, the neuro-symbolic literature, and SECI-style organizational knowledge theory. The added value is the synthesis, the framing around multi-agent and cross-generation accumulation, and the operational invariants. The authors engage honestly with related positions (Chen's large knowledge model, LKM for science).

Methodological Rigor

There are no experiments, so rigor is a matter of argumentation. The paper is careful in several ways:

It scopes its bottleneck hypothesis to large, heterogeneous, multi-contributor, changing stores.

It concedes that a strong LLM with raw records may match accuracy, and locates the benefit in reuse, auditability and continuity instead.

It includes an "Alternative Views" section with fairly strong steelmen.

It labels Eq. 2 as non-optimizable and as a statement of purpose only.

It proposes a falsifiable comparison.

Weaknesses:

The central claim, that internalization is the bottleneck and a persistent KFM is worthwhile, is not backed by any pilot evidence. The paper itself says the increment might appear only in some regimes.

The KFM is defined by desiderata and is only loosely specified architecturally.

Eq. 2 is hard to operationalize, because description length, the distribution Q and the utility u are undefined.

The maturity assessments (Tables 3 and 5) are subjective, self-described as "our assessment".

There is heavy self-citation to the authors' group (AutoSchemaKG, ASER, CAT/CANDLE, KGPFN, etc.), and some reference dates look anomalous. The arXiv ID and date in the header are odd.

Some claims of absence, such as "no general method extracts rules from tabular FMs", are hard to verify.

Potential Impact

The paper could serve as an organizing reference for work on agent memory, knowledge-graph construction, and neural graph databases. It offers a vocabulary: internalization, condition-preserving conceptualization, the invariants, and KFM. Several elements could become concrete research programs or benchmarks:

Invariant-based evaluation of memory and KG operators, as opposed to end-task accuracy alone.

Extraction of explicit rules from tabular and time-series foundation models.

Knowledge-conditioning interfaces for non-language foundation models (Gap 3).

The raw-records-agent versus knowledge-layer comparison.

Enterprise, consumer-agent and scientific-discovery communities could draw on it. Its influence will depend on whether the community builds benchmarks around it. Position papers of this kind often get cited at moderate levels, especially when they are tied to a research group's broader program, as this one is.

Timeliness & Relevance

Timeliness is high. Agent memory, self-evolving agents, skill libraries, context engineering and knowledge conflicts are all active areas, and agent turnover across model generations is a real concern. The explicit-versus-parametric knowledge debate and the question of whether long context or scale makes structured stores obsolete are live. Treating knowledge portability across model generations and governance concerns (deletion propagation, permissions, auditability) as first-class matters is a useful framing.

Strengths

Coherent and well-organized argument with a clear thesis, principles P1–P3 and falsifiable commitments.

Concrete, checkable invariants that go beyond generic "use a KG" advice.

Broad, well-organized literature mapping across four knowledge types (factual, conceptual, procedural, metacognitive), the carrier forms, and the use channels.

Useful distinctions: observation vs. experience vs. model judgment; knowledge about the environment vs. advice tied to an old agent; fidelity to a model vs. recovery of the true mechanism.

Honest treatment of the counterarguments and costs.

Limitations

No empirical validation of any claim, including the bottleneck hypothesis.

Vision-heavy. KFM remains an aspiration, and "foundation model" framing risks being a label.

Scope is very wide (seven challenges), so each is shallow.

Self-citation and an in-group perspective (eventuality and conceptualization lines) may skew emphasis.

Heavily text- and KG-oriented in its details despite claiming modality-agnosticism.

Dense prose in places, with notation (T, R, U, K, κ) that is useful but not developed into formal results.

Practical obstacles are underplayed: incentives for sharing across organizations, trust and adversarial contributions, and governance of a shared layer.

Overall this is a thoughtful, timely position paper with a useful vocabulary and a credible agenda. It is not a conceptual breakthrough, and its impact depends on follow-up empirical work.

Rating:6/ 10

Significance 6Rigor 5Novelty 5.5Clarity 7.5

Generated Nov 3, 2026

Comparison History (0)

No comparisons yet.

Scorecard· 12/16

6.0/10 impact

Well-argued, timely position paper with concrete invariants and a falsifiable agenda, but no empirical validation and substantial overlap with existing KG/agent-memory/large-knowledge-model framings.