Kumiho: Graph-Based Cognitive Memory System

• Kumiho is a graph-native cognitive memory and asset management system that records beliefs as immutable revisions and supports provenance tracking.
• It employs a property-graph structure with mutable tags and typed dependency edges to merge cognitive state management with artifact versioning based on formal AGM semantics.
• Benchmark results demonstrate near-perfect adversarial accuracy and robust performance, underlining its practical impact on multi-agent pipelines and audit-ready collaboration.

Kumiho is a graph-native cognitive-memory and asset-management architecture for AI agents that unifies belief tracking, provenance, revision control, and work-product management in a single property graph (Park, 18 Mar 2026). Its central claim is that the structural primitives required for cognitive memory—immutable revisions, mutable tag pointers, typed dependency edges, and URI-based addressing—are the same primitives required to manage agent-produced artifacts as versionable assets. On that basis, Kumiho stores both beliefs and work products in one graph-native system, while grounding memory operations in formal belief revision semantics derived from the AGM framework and Hansson’s belief-base postulates (Park, 18 Mar 2026).

1. Conceptual scope and unifying design

Kumiho is designed to give AI agents “true cognitive memory”: the ability to record each belief or decision as an immutable snapshot, track how it changes over time, recall its provenance, and automatically propagate updates through dependent conclusions (Park, 18 Mar 2026). The same graph structure simultaneously versions and links work products such as code commits, images, and documents, so that multi-agent pipelines become first-class citizens of the memory graph.

The architecture’s core insight is that memory and artifact management need not be treated as separate systems. In Kumiho, the same primitives support both domains: immutable revisions preserve historical state, mutable tags determine the active retrieval surface, typed edges capture provenance and dependency structure, and URI-based addressing makes every memory or artifact stably referenceable across sessions, agents, and human inspection (Park, 18 Mar 2026). This suggests a unification of cognitive state management and operational asset lineage within a single storage substrate rather than a layered composition of separate memory and repository systems.

A plausible implication is that Kumiho is not merely a retrieval layer over conversational history. It is a versioned graph in which beliefs, summaries, artifacts, and dependency structure are co-managed under one revision discipline. That interpretation is supported by the paper’s emphasis on auditability, durable versioning, and explicit propagation through dependent conclusions (Park, 18 Mar 2026).

2. Property-graph representation and addressing model

In Kumiho’s property graph, each conceptual “Item” has an ordered chain of immutable “Revision” nodes, with each revision carrying structured metadata such as summary, topics, keywords, and optionally an embedding vector (Park, 18 Mar 2026). A small set of mutable “Tags,” including examples such as “current” or “approved,” points to exactly one revision per item. This yields a finite belief base

$$\mathcal{B}(\tau)\;=\;\bigcup_{t\in\mathrm{dom}(\tau)}\varphi\bigl(\tau(t)\bigr),$$

where $\tau$ is the tag-to-revision mapping and $\varphi(r)$ is the revision’s propositional content (Park, 18 Mar 2026).

Typed, directed edges encode distinct semantic roles. The architecture specifies edges such as Depends_On, Derived_From, Supersedes, Referenced, Contains, and Created_From, which express evidential provenance, validity dependencies, revision chains, and generative lineage (Park, 18 Mar 2026). This edge vocabulary is central to the claim that the graph is “native” rather than merely graph-backed: provenance and dependency structure are not auxiliary annotations but part of the architecture’s operational semantics.

Every memory or artifact is addressable by a stable URI of the form

$$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$

which supports cross-session, cross-agent, and human-auditable references (Park, 18 Mar 2026). The addressing model matters because Kumiho treats both beliefs and assets as durable graph objects rather than ephemeral context fragments. This, in turn, enables explicit references to past revisions instead of reconstructing state indirectly from transcripts.

3. Belief revision semantics and formal guarantees

Kumiho maps its memory operations to the AGM framework for rational belief change, but does so at the belief-base level, following Hansson’s treatment of stored beliefs that need not be deductively closed (Park, 18 Mar 2026). The paper defines three key operations.

Expansion adds a proposition $A$ by creating a new revision:

$\mathcal{B}+A.$

Contraction removes any tag-referenced revision whose content explicitly contains $A$, while marking the underlying item deprecated and never deleting history:

$\mathcal{B}\div A.$

Revision is defined by the Levi identity,

$$\mathcal{B}*A \equiv \text{first contract } \neg A,\text{ then expand } A,$$

but is implemented atomically by creating a new revision $r'$ with $\tau$0, linking it by a Supersedes edge to the prior tag-referenced revision, and repointing the tag (Park, 18 Mar 2026).

The architecture proves satisfaction of the basic AGM postulates $\tau$1–$\tau$2 and Hansson’s Relevance and Core-Retainment by construction (Park, 18 Mar 2026). In the paper’s formulation:

• Success ($\tau$3): after $\tau$4, $\tau$5.
• Inclusion ($\tau$6): $\tau$7.
• Vacuity ($\tau$8): if $\tau$9 is consistent with $\varphi(r)$0, no retraction occurs beyond adding $\varphi(r)$1.
• Consistency ($\varphi(r)$2): the new revision replaces the old one, so contradictory beliefs do not co-exist in the active tag set.
• Extensionality ($\varphi(r)$3): in Kumiho’s propositional logic of ground triples, logical equivalence reduces to syntactic identity, so identical beliefs yield identical revisions.
• Relevance (Hansson): during contraction, only revisions containing $\varphi(r)$4 are detagged.
• Core-Retainment (Hansson): every removed belief co-occurred with $\varphi(r)$5 in the same revision (Park, 18 Mar 2026).

A significant non-classical feature is that Recovery is intentionally violated. If a revision contains $\varphi(r)$6, contracting $\varphi(r)$7 archives both, and re-expanding creates $\varphi(r)$8 only rather than restoring $\varphi(r)$9 automatically (Park, 18 Mar 2026). This is not presented as a defect but as a consequence of immutable revision history and auditable rollback via explicit tag reassignment. In that sense, Kumiho prioritizes transparent state transition over implicit restoration semantics.

The paper further states that supplementary postulates $$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$0 and $$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$1 can be argued in many cases, but that a full AGM representation theorem would require an explicit entrenchment ordering; this remains an open formal direction (Park, 18 Mar 2026). That limitation is important because it locates Kumiho’s formal contribution not as a complete reconstruction of AGM, but as a graph-operational semantics satisfying a substantial subset of belief-revision requirements.

4. Dual-store implementation and retrieval pipeline

Kumiho uses a dual-store model consisting of Redis for working memory and Neo4j for long-term graph storage (Park, 18 Mar 2026). Working memory is a Redis buffer with sub-10 ms in-process read/write, holding the current conversational turn buffer with TTL. Long-term memory is a Neo4j property graph storing items, revisions, tags, edges, and deprecation flags. Raw conversation transcripts and large artifacts remain local under a BYO-storage model; the graph stores metadata, summaries, pointers, and embeddings rather than full payloads (Park, 18 Mar 2026).

Retrieval is hybrid. Each query executes in Neo4j as a two-branch search: full-text BM25 over summary, topics, keywords, and title, and vector similarity using cosine over 1 536-dim embeddings (Park, 18 Mar 2026). The resulting scores $$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$2 and $$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$3 are fused as

$$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$4

with type-aware weight $$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$5 (Park, 18 Mar 2026). Deprecated items are excluded at query time by a mandatory WHERE [NOT](https://www.emergentmind.com/topics/neural-organ-transplantation-not) item.deprecated clause, which the paper identifies as guaranteeing retrieval consistency under $$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$6.

The architecture also exposes graph-traversal tools—AnalyzeImpact, ShortestPath, and TraverseEdges—for causal or dependency reasoning (Park, 18 Mar 2026). These tools are operationally distinct from surface retrieval, indicating that Kumiho separates candidate recall from explicit graph reasoning.

Three architectural innovations are presented as central to performance (Park, 18 Mar 2026):

Innovation	Mechanism	Stated role
Prospective indexing	LLM generates 3–5 hypothetical future scenarios during session consolidation	Supports retrieval when query phrasing differs from the original conversation
Event extraction	LLM extracts structured causal events as <description → consequence> pairs	Preserves fine-grained causal details often omitted by plain narrative summaries
Client-side LLM reranking	Embedding pre-filter with cosine $$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$7, then the consuming agent’s LLM selects the most relevant sibling revision	Provides zero-cost reranking and improves automatically as the answer model improves

Prospective indexing extends the indexed text with hypothetical future implications, thereby addressing semantic disconnect between original memory encoding and later query form (Park, 18 Mar 2026). Event extraction preserves causal structure in summaries rather than leaving it latent in narrative text. Client-side reranking is noteworthy because it is described as memory-infrastructure-agnostic: the same answer model already being invoked for final response generation chooses among sibling revisions after a lightweight embedding pre-filter (Park, 18 Mar 2026).

5. Benchmark results

Kumiho is evaluated on both LoCoMo and LoCoMo-Plus (Park, 18 Mar 2026). On LoCoMo, using the official token-level F1 with Porter stemming metric, it achieves 0.447 four-category F1 over $$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$8 factual retrieval questions and 0.565 overall F1 when adversarial refusal accuracy is included, with adversarial refusal accuracy of 97.5% over $$\mathrm{kref://project/space/item.kind?r=N\,[{data}a=artifact]},$$9 questions (Park, 18 Mar 2026). The paper attributes the near-perfect adversarial result to its belief-revision semantics: the graph never contains fabricated beliefs, so there is no fabricated content available for retrieval-conditioned hallucination.

On LoCoMo-Plus, which is described as a Level-2 cognitive memory benchmark testing implicit constraint recall, Kumiho achieves 93.3% judge accuracy on 401 implicit-constraint queries with GPT-4o as the answer model, 88% with GPT-4o-mini, and 98.5% recall accuracy for retrieving at least one relevant memory (Park, 18 Mar 2026). The paper reports that this exceeds the best published baseline, Gemini 2.5 Pro at 45.7%, by nearly 48 percentage points. It also states that independent reproduction by the benchmark authors yielded results in the mid-80% range, still substantially outperforming all published baselines (Park, 18 Mar 2026).

Three empirical findings are emphasized (Park, 18 Mar 2026). First, enrichment matters: a pre-enrichment baseline on 200 entries scored 61.6%, while adding event extraction, prospective indexing, and client-side reranking raised accuracy to 93.3% and removed the long-horizon cliff for gaps greater than 6 months, from 37.5% to 84.4%. Second, retrieval is model-decoupled in the sense that recall accuracy is invariant across GPT-4o-mini and GPT-4o, while end-to-end answer accuracy improves by +5.3 percentage points when only the answer model is upgraded. Third, the evaluation strategy is cost efficient: bulk summarization, query reformulation, and consolidations run on GPT-4o-mini, only the final answer uses GPT-4o, and the total cost for 401 entries is approximately \$14 (Park, 18 Mar 2026).

These results support the paper’s distinction between retrieval quality and answer-model quality. Retrieval recall is treated as a property of the architecture, whereas final judged accuracy depends partly on the downstream model consuming the retrieved context (Park, 18 Mar 2026). That separation is a recurring theme in the system design.

6. Significance, applications, and open directions

Kumiho’s principal significance lies in the combination of formal belief revision guarantees with a practical memory-and-artifact architecture (Park, 18 Mar 2026). By satisfying $A$0–$A$1 together with Relevance and Core-Retainment, and by enforcing immutable revisions plus mutable tags, it aims to guarantee minimal-change belief revision, absence of hidden side effects, and a retrieval surface in which contradictory beliefs do not co-exist (Park, 18 Mar 2026). The paper explicitly connects these properties to consistency, trust, auditability, and production robustness.

Potential applications listed in the paper include fully autonomous multi-agent pipelines in which each agent’s input and output artifacts are versioned and linked in the same cognitive graph; human–AI collaboration with audit-ready memory traces in regulated domains such as finance and healthcare; and longitudinal analysis of evolving preferences or system configurations over months or years (Park, 18 Mar 2026). These use cases follow directly from the architecture’s stable addressing, explicit provenance, and long-term versioning.

The paper also identifies several open directions (Park, 18 Mar 2026). These include formalizing an entrenchment ordering to prove $A$2 and $A$3 or to develop partial-merge belief operators based on Konieczny–Pino Pérez; extending the propositional graph logic toward AGM-compatible fragments of richer knowledge-representation languages; running large-scale ablations of retrieval fusion strategies such as RRF versus max fusion and tuning vector/fulltext calibration; integrating predictive consolidation following Letta’s sleep-time compute approach; and incorporating temporal recency as a native retrieval signal to prioritize newer revisions.

A common misconception would be to treat Kumiho as solely a long-context retrieval system. The paper’s own framing is narrower and more formal: it is a graph-native, versioned memory architecture with explicit operational correspondence to belief revision semantics, and with artifact management folded into the same graph substrate (Park, 18 Mar 2026). Another possible misconception would be to assume that formal semantics imply complete AGM compliance; the text is explicit that Recovery is intentionally violated and that a full treatment of supplementary AGM postulates remains open (Park, 18 Mar 2026). Those caveats are central to understanding Kumiho’s position: it is formally grounded, but not a maximalist realization of every classical postulate.