{"@id":"https://github.com/shimo4228/agent-knowledge-cycle#knowledge-graph","@type":["Dataset","CreativeWork"],"name":"Agent Knowledge Cycle (AKC) Knowledge Graph","description":"Canonical machine-readable relationship map for the Agent Knowledge Cycle line. Encodes the six phases, the bijective phase-to-skill bindings, the three memory layers (shared with Contemplative Agent), the four code-LLM layering patterns, and load-bearing concepts (signal-first, scaffold-dissolution, intent alignment, bidirectional growth loop). AI agents and LLM-based search systems should read this graph before summarizing the line or following individual document links.","isBasedOn":"https://github.com/shimo4228/agent-knowledge-cycle","mainEntity":"https://doi.org/10.5281/zenodo.19200726"}
{"@id":"https://doi.org/10.5281/zenodo.19200726","@type":["ResearchLine","ScholarlyArticle"],"name":"Agent Knowledge Cycle (AKC)","alternateName":[{"@value":"Agent Knowledge Cycle","@language":"en"},{"@value":"エージェント知識サイクル","@language":"ja"}],"description":"Six-phase bidirectional growth loop in which agent behavior and the operator's judgment co-develop over time, sustaining intent alignment that tests cannot check on their own. Three stacked layers — principles (ADRs), patterns (design-pattern skills), and implementation (composable skills) — decouple rate of change. Refers to a mechanism, not a disposition or a practice.","identifier":"10.5281/zenodo.19200726","url":"https://github.com/shimo4228/agent-knowledge-cycle","author":"https://orcid.org/0009-0002-6168-4162","siblingOf":["https://doi.org/10.5281/zenodo.19212118","https://doi.org/10.5281/zenodo.19652013"],"derivesFrom":"https://doi.org/10.5281/zenodo.19212118","definesConcept":["https://shimo4228.github.io/shimo4228/vocab#concept/six-phase-loop","https://shimo4228.github.io/shimo4228/vocab#concept/three-layer-structure","https://shimo4228.github.io/shimo4228/vocab#concept/scaffold-dissolution","https://shimo4228.github.io/shimo4228/vocab#akc/concept/signal-first","https://shimo4228.github.io/shimo4228/vocab#akc/concept/intent-alignment","https://shimo4228.github.io/shimo4228/vocab#akc/concept/harness-alignment","https://shimo4228.github.io/shimo4228/vocab#akc/concept/harness-drift","https://shimo4228.github.io/shimo4228/vocab#akc/concept/bidirectional-growth-loop","https://shimo4228.github.io/shimo4228/vocab#akc/concept/loop-failure-modes","https://shimo4228.github.io/shimo4228/vocab#akc/concept/self-reingestion","https://shimo4228.github.io/shimo4228/vocab#akc/concept/observed-vs-generated","https://shimo4228.github.io/shimo4228/vocab#akc/concept/two-stage-distill","https://shimo4228.github.io/shimo4228/vocab#akc/concept/human-approval-gate","https://shimo4228.github.io/shimo4228/vocab#akc/concept/genre-neutral","https://shimo4228.github.io/shimo4228/vocab#akc/concept/cognitive-economy","https://shimo4228.github.io/shimo4228/vocab#akc/concept/code-llm-layering"],"contrastsWith":["https://shimo4228.github.io/shimo4228/vocab#prior-art/coala","https://shimo4228.github.io/shimo4228/vocab#prior-art/voyager","https://shimo4228.github.io/shimo4228/vocab#prior-art/generative-agents","https://shimo4228.github.io/shimo4228/vocab#prior-art/memgpt","https://shimo4228.github.io/shimo4228/vocab#prior-art/reme","https://shimo4228.github.io/shimo4228/vocab#prior-art/agent-workflow-memory","https://shimo4228.github.io/shimo4228/vocab#prior-art/meta-harness","https://shimo4228.github.io/shimo4228/vocab#prior-art/agent-drift"]}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/research","@type":["Phase","DefinedTerm"],"name":"Research phase","alternateName":[{"@value":"Research","@language":"en"},{"@value":"Research（探索）","@language":"ja"}],"description":"First of six AKC phases. Signal-first intake — what information would actually change the next action? Bound bijectively to the search-first skill.","ordinal":1}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/extract","@type":["Phase","DefinedTerm"],"name":"Extract phase","alternateName":[{"@value":"Extract","@language":"en"},{"@value":"Extract（抽出）","@language":"ja"}],"description":"Second of six AKC phases. Capture reusable patterns from sessions with quality gates before they are saved. Bound bijectively to the learn-eval skill.","ordinal":2}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/curate","@type":["Phase","DefinedTerm"],"name":"Curate phase","alternateName":[{"@value":"Curate","@language":"en"},{"@value":"Curate（選別）","@language":"ja"}],"description":"Third of six AKC phases. Audit accumulated skills and rules for staleness, conflicts, and redundancy. Bound bijectively to the skill-stocktake skill.","ordinal":3}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/promote","@type":["Phase","DefinedTerm"],"name":"Promote phase","alternateName":[{"@value":"Promote","@language":"en"},{"@value":"Promote（昇格）","@language":"ja"}],"description":"Fourth of six AKC phases. Distill cross-cutting principles whose recurrence across skills, memory, and conversation has itself become the signal into rules — a holistic judgment, not a numeric threshold. Bound bijectively to the rules-distill skill.","ordinal":4}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/measure","@type":["Phase","DefinedTerm"],"name":"Measure phase","alternateName":[{"@value":"Measure","@language":"en"},{"@value":"Measure（測定）","@language":"ja"}],"description":"Fifth of six AKC phases. Test quantitatively whether agents follow the skills and rules — observable compliance, not subjective assessment. The observable surface includes agent text, not only tool calls: thinking-centric phases (Research, Curate) carry their compliance in reasoning, verdicts, and plans, so a Measure instrument must promote agent text to observable events before scoring or it systematically under-reports those phases (ADR-0016). Bound bijectively to the skill-comply skill.","ordinal":5}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/maintain","@type":["Phase","DefinedTerm"],"name":"Maintain phase","alternateName":[{"@value":"Maintain","@language":"en"},{"@value":"Maintain（保守）","@language":"ja"}],"description":"Sixth of six AKC phases. Audit documentation roles for overlap and stale content; keep CLAUDE.md / CODEMAPS / ADR / README clean. Bound bijectively to the context-sync skill.","ordinal":6}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#memory-layer/episode-log","@type":["MemoryLayer","DefinedTerm"],"name":"Episode Log (Layer 1)","alternateName":[{"@value":"Episode Log","@language":"en"},{"@value":"エピソードログ","@language":"ja"}],"description":"Layer 1 of the three-layer memory architecture (ADR-0003): raw, immutable JSONL episode logs. The owner-only source-of-truth from which higher layers are distilled.","level":1}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#memory-layer/knowledge","@type":["MemoryLayer","DefinedTerm"],"name":"Knowledge (Layer 2)","alternateName":[{"@value":"Knowledge Layer","@language":"en"},{"@value":"知識層","@language":"ja"}],"description":"Layer 2 of the three-layer memory architecture (ADR-0003): distilled patterns, time-decayed and forbidden-substring validated. Promoted from Layer 1; gated when promoted further to Layer 3.","level":2}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#memory-layer/identity","@type":["MemoryLayer","DefinedTerm"],"name":"Identity / Rules (Layer 3)","alternateName":[{"@value":"Identity Layer","@language":"en"},{"@value":"アイデンティティ層","@language":"ja"}],"description":"Layer 3 of the three-layer memory architecture (ADR-0003): persona, ranked skills, cross-cutting rules. Deterministic. Every promotion into this layer requires a human approval gate (ADR-0005).","level":3}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0002-immutable-episode-log.md","@type":["ADR","TechArticle"],"name":"ADR-0002: Immutable Episode Log as Source of Truth","description":"The raw episode log is append-only and immutable; every higher layer is a derivation. Treats accumulated state as the auditable trail rather than as authority."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0003-three-layer-distillation.md","@type":["ADR","TechArticle"],"name":"ADR-0003: Three-Layer Distillation (Raw → Knowledge → Identity / Rules)","description":"The memory architecture: Layer 1 raw, Layer 2 knowledge, Layer 3 identity / rules. Decouples rate of change so principles stay stable while implementations evolve independently."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0004-two-stage-distill-pipeline.md","@type":["ADR","TechArticle"],"name":"ADR-0004: Two-Stage Distill Pipeline (Free-form → Format)","description":"Free-form reasoning followed by structured formatting. Decouples thinking from formatting so the model is not asked to do both simultaneously."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0005-human-approval-gate.md","@type":["ADR","TechArticle"],"name":"ADR-0005: Human Approval Gate for Behavior-Modifying Changes","description":"Promotions that produce behavior-modifying writes require named human sign-off. The cycle is not autonomous; the gate is structural."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0008-code-and-llm-collaboration.md","@type":["ADR","TechArticle"],"name":"ADR-0008: Code and LLM Collaboration","description":"Four code-LLM layering patterns — guard, filter, judge, orchestrator — formalize when to use code, when to use the LLM, and how to layer them. The patterns recur across all six phases."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0009-akc-is-a-cycle-not-a-harness.md","@type":["ADR","TechArticle"],"name":"ADR-0009: AKC is a Cycle, Not a Harness","description":"AKC is the cycle — the mechanism — not the substrate. Distinguishes AKC from harness frameworks; the cycle is portable across harnesses, the harness is not the cycle."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0010-human-cognitive-resource-as-central-constraint.md","@type":["ADR","TechArticle"],"name":"ADR-0010: Human Cognitive Resource as Central Constraint","description":"Human attention and judgment, not compute or context, is the central constraint that does not scale with the model. Every other framework optimizes the agent side; AKC asks the inverse question."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0011-cycle-applies-to-any-knowledge-body.md","@type":["ADR","TechArticle"],"name":"ADR-0011: Cycle Applies to Any Knowledge Body (Genre-Neutral)","description":"The cycle applies to behavioral patterns, domain expertise, and constitutional values alike — not specific to one genre of knowledge. Genre-neutrality is structural, not aspirational."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0012-front-load-three-core-themes.md","@type":["ADR","TechArticle"],"name":"ADR-0012: Front-load the Three Core Themes in Front-door Documentation","description":"Three themes — human attention as constraint, intent alignment over correctness, bidirectional growth — must appear in the project's reading surface before deeper material. Reading-order discipline."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0013-positioning-within-agent-memory-literature.md","@type":["ADR","TechArticle"],"name":"ADR-0013: Positioning Within the Agent-Memory Literature","description":"Explicit Related-Work positioning. Concedes that AKC's operations (skill induction, memory refinement, reflection) are not novel in isolation — Voyager, Agent Workflow Memory, ReMe, LangMem, Generative Agents, MemGPT, and CoALA name each one — then locates AKC's delta on the shared axis: a structural human approval gate where the prior art runs autonomously, a bidirectional human-judgment target where the prior art optimizes the agent, and human attention as the scarce resource where the prior art treats agent capability or context as binding. Orthogonal to ADR-0009's harness positioning."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0014-failure-modes-of-the-bidirectional-loop.md","@type":["ADR","TechArticle"],"name":"ADR-0014: Failure Modes of the Bidirectional Loop","description":"Theme 3's failure twin. A loop that can grow human judgment can also erode it. Names three mechanism-level failure modes — gate complacency, deskilling, delegation-feedback divergence — and the structural defenses already in the cycle: the approval gate as circuit-breaker, Curate and Promote as active-judgment acts, and the gate's diff-and-commit operation as forced articulation. Experimental."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0015-loop-failure-modes-self-reingestion.md","@type":["ADR","TechArticle"],"name":"ADR-0015: Self-Reingestion — When the Cycle Feeds on Its Own Output","description":"The cycle feeds on its own output: promoted rules load into every session and shape the records, and agent-written notes are read again by the next distillation. As the self-generated share of the input grows, the agent's phrasings echo back as if they were facts and each re-summarization weakens the link to the original events — a cycle designed to prevent knowledge decay can become the decay vector. Distilled from a production observation in the Contemplative Agent substrate. Guards: only the record of what happened is observed ground, the self-generated share of promoted artifacts is kept visible (Measure), and the Human Approval Gate is the circuit-breaker. Experimental."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0016-measuring-thinking-centric-phases.md","@type":["ADR","TechArticle"],"name":"ADR-0016: Measuring Thinking-Centric Phases","description":"A Measure instrument that observes only tool calls systematically under-reports the thinking-centric phases (Research, Curate) whose load-bearing behavior lives in agent text — reasoning, verdicts, plans — not in side effects. Records the falsifiable claim and the corrective: a Measure instrument must promote agent text to observable events before scoring. Ties the Measure phase to Theme 2 (intent alignment lives in reasoning; correctness lives in tool calls)."}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0017-harness-alignment-and-drift.md","@type":["ADR","TechArticle"],"name":"ADR-0017: Harness Alignment and Harness Drift","description":"Establishes two terms by derivation from established literature rather than fresh coinage — the second positioning ADR after ADR-0013, applied to the software-evolution and alignment literatures. Harness alignment: the continuous, human-gated activity of keeping an agent's harness (skills, rules, prompts, documentation) aligned with the operator's evolving intent — extends intent alignment (Christiano 2018) to the configuration layer and across time (cf. Lehman's Law of Continuing Change, 1980). Harness drift: the failure mode, in lineage with architectural drift (Perry & Wolf 1992), practical drift (Snook 2000), and agent drift (arXiv:2601.04170). Contrast: autonomous, score-driven harness optimization (Meta-Harness, arXiv:2603.28052). Also records the verified vocabulary gap: the 2026 agent-drift literature does not cite the classical software-evolution lineage."}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#concept/six-phase-loop","@type":["Concept","DefinedTerm"],"name":"six-phase loop","description":"Bidirectional growth loop running Research, Extract, Curate, Promote, Measure, Maintain. Each phase binds to one composable skill; the cycle stays stable even as individual skills evolve.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0009-akc-is-a-cycle-not-a-harness.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#concept/three-layer-structure","@type":["Concept","DefinedTerm"],"name":"three-layer structure","description":"AKC's stacked architecture: principles (ADRs), patterns (design-pattern skills), implementation (composable skills). Separating layers decouples rate of change.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0003-three-layer-distillation.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#concept/scaffold-dissolution","@type":["Concept","DefinedTerm"],"name":"scaffold dissolution","description":"Property that explicit AKC skills can be dropped once the cycle has been internalized. The skills are scaffolding, not the goal. Dissolution is the intended end state, not a fallback.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0009-akc-is-a-cycle-not-a-harness.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/signal-first","@type":["Concept","DefinedTerm"],"name":"signal-first","description":"Research-phase intake principle: what information would actually change the next action? Anything outside that is out of scope. Search widely, intake narrowly. (ADR-0010 derivative.)","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0010-human-cognitive-resource-as-central-constraint.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/intent-alignment","@type":["Concept","DefinedTerm"],"name":"intent alignment","description":"Sustaining alignment between agent behavior and the operator's evolving intent across sessions, distinguished from per-output correctness. Intent itself moves as judgment sharpens through use; correctness can be checked by tests, alignment cannot.","groundedIn":["https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0010-human-cognitive-resource-as-central-constraint.md","https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0012-front-load-three-core-themes.md"]}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/harness-alignment","@type":["Concept","DefinedTerm"],"name":"harness alignment","description":"The continuous, human-gated activity of keeping an agent's harness — its configuration layer: skills, rules, prompts, documentation — aligned with the operator's evolving intent. Extends intent alignment (Christiano 2018) from agent behavior to the artifacts that shape behavior, and across time: alignment is sustained through a cycle, not configured once (cf. Lehman's Law of Continuing Change, 1980). The alignment target is operator intent, not model values. The contrast is harness optimization (Meta-Harness, arXiv:2603.28052): autonomous, score-driven improvement on the correctness axis. AKC's six phases operationalize harness alignment: Measure and Maintain detect harness drift; Curate and Promote correct it through the human approval gate.","groundedIn":["https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0017-harness-alignment-and-drift.md","https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0009-akc-is-a-cycle-not-a-harness.md"],"extends":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/intent-alignment","hasFailureMode":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/harness-drift","contrastsWith":"https://shimo4228.github.io/shimo4228/vocab#prior-art/meta-harness"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/harness-drift","@type":["Concept","DefinedTerm"],"name":"harness drift","description":"Harness alignment's failure mode: the gradual uncoupling of the harness from operator intent when the cycle does not run — skills go stale, rules stop matching practice, documentation diverges from code. Named in lineage with architectural drift (Perry & Wolf 1992: divergence by insensitivity), practical drift (Snook 2000: practice uncoupling from written procedure), and agent drift (arXiv:2601.04170: behavioral-level deviation from original intent). An artifact-side failure, distinct from the human-side loop failure modes of ADR-0014 (gate complacency, deskilling, delegation-feedback divergence); the two compound but are recorded separately.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0017-harness-alignment-and-drift.md","derivesFrom":["https://shimo4228.github.io/shimo4228/vocab#prior-art/architectural-drift-perry-wolf","https://shimo4228.github.io/shimo4228/vocab#prior-art/practical-drift-snook","https://shimo4228.github.io/shimo4228/vocab#prior-art/agent-drift"]}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/bidirectional-growth-loop","@type":["Concept","DefinedTerm"],"name":"bidirectional growth loop","description":"Agent behavior and human judgment co-develop. As the human curates and promotes knowledge, their judgment about what makes good agent behavior also sharpens. Not a one-directional optimization loop.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0012-front-load-three-core-themes.md","hasFailureMode":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/loop-failure-modes"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/loop-failure-modes","@type":["Concept","DefinedTerm"],"name":"loop failure modes","description":"The failure twin of the bidirectional growth loop. A loop that can grow human judgment can also erode it. Three mechanism-level failure modes: gate complacency (a reliable agent trains default-approval, thinning the judgment behind the Promote gate), deskilling (over-delegation starves the supervisory faculty that review depends on), and delegation-feedback divergence (delegation continues while the coupling that would correct the agent breaks, so the loop still produces output no human is steering). The cycle's existing structure resists, but does not immunize against, these modes.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0014-failure-modes-of-the-bidirectional-loop.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/self-reingestion","@type":["Concept","DefinedTerm"],"name":"self-reingestion","description":"A failure mode of the cycle: the agent's own output re-enters as input — promoted rules load into every session and shape what the records contain, and agent-written notes and summaries are read again by the next distillation. As the self-generated share of the input grows, two degradations follow: echo (the agent's own phrasings, re-read across cycles, are eventually promoted as if they were observed facts) and grounding loss (each re-summarization weakens the link to the original events). The strength of the effect tracks the self-generated share of the input — installs with a human in every session dilute it, autonomous self-narrative-heavy installs concentrate it. Observed in production in the Contemplative Agent substrate; distilled to the mechanism level by ADR-0015. Guards: only the record of what happened is observed ground, the self-generated share is kept visible (Measure), and the Human Approval Gate breaks the loop before generated content reaches the always-loaded layer.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0015-loop-failure-modes-self-reingestion.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/observed-vs-generated","@type":["Concept","DefinedTerm"],"name":"observed-vs-generated composition signal","description":"A Measure-phase observability metric introduced by ADR-0015: the share of a promoted artifact's input that is the agent's own earlier output versus observed record. It is an observability signal, not an intervention — it does not change retrieval, ranking, or promotion. Its purpose is to make self-reingestion visible so a rising self-generated share can be noticed before it compounds. The encoding (per-artifact counts, sampled audit) is left to the harness.","groundedIn":["https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0015-loop-failure-modes-self-reingestion.md","https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0002-immutable-episode-log.md"]}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/two-stage-distill","@type":["Concept","DefinedTerm"],"name":"two-stage distill pipeline","description":"Free-form reasoning followed by structured formatting. Decouples thinking from formatting so the model is not asked to do both simultaneously. Codified as ADR-0004.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0004-two-stage-distill-pipeline.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/human-approval-gate","@type":["Concept","DefinedTerm"],"name":"human approval gate","description":"Structural checkpoint for behavior-modifying writes. AKC is not autonomous; promotion across layers requires named human sign-off. Codified as ADR-0005.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0005-human-approval-gate.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/genre-neutral","@type":["Concept","DefinedTerm"],"name":"genre-neutral","description":"The cycle applies to behavioral patterns, domain expertise, and constitutional values alike — not specific to one genre of knowledge. Codified as ADR-0011.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0011-cycle-applies-to-any-knowledge-body.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/cognitive-economy","@type":["Concept","DefinedTerm"],"name":"cognitive economy","description":"Treats human attention and judgment as the scarce resource. Information that does not change an action does not deserve to be held; intake is where attention is spent.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0010-human-cognitive-resource-as-central-constraint.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/concept/code-llm-layering","@type":["Concept","DefinedTerm"],"name":"code-LLM layering","description":"Four patterns formalize when to use code, when to use the LLM, and how to layer them: guard (code validates LLM output against schema, forbidden-pattern, and size checks after generation, before it touches persistent state), filter (code narrows a large, noisy input before the LLM is invoked), judge (LLM decides among bounded options provided by code, which enforces the deterministic action), orchestrator (code owns the loop — termination, retries, state — and the LLM is the worker doing one semantic task per step). Codified as ADR-0008.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0008-code-and-llm-collaboration.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/pattern/guard","@type":["Concept","DefinedTerm"],"name":"code-LLM pattern: guard","description":"Post-LLM output validation. Code validates LLM output against schema, forbidden-pattern, and size checks after generation, before it touches persistent state. The LLM produces structured output; code is the gate it must pass before any durable write.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0008-code-and-llm-collaboration.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/pattern/filter","@type":["Concept","DefinedTerm"],"name":"code-LLM pattern: filter","description":"Pre-LLM input narrowing. Code narrows a large, noisy input first — by date, type, size, or structural relevance — before the LLM is invoked, and the LLM does semantic work on the curated subset. Used when the filtering criteria are mechanical but the accept-criteria are semantic.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0008-code-and-llm-collaboration.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/pattern/judge","@type":["Concept","DefinedTerm"],"name":"code-LLM pattern: judge","description":"LLM judge + code enforce. The LLM decides among bounded, named options provided by code (accept / reject / score); code then takes the deterministic action, with a human gate for high-stakes changes (ADR-0005). Used when the decision is semantic but the option space is enumerable and the consequent action must stay code-owned.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0008-code-and-llm-collaboration.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#akc/pattern/orchestrator","@type":["Concept","DefinedTerm"],"name":"code-LLM pattern: orchestrator","description":"Code owns the loop; the LLM is the worker. A deterministic loop owns termination, retries, state, and error handling, and the LLM does one focused semantic task per step. The LLM never owns termination or durable state. Never invert this.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0008-code-and-llm-collaboration.md"}
{"@id":"https://github.com/shimo4228/claude-skill-search-first","@type":["EcosystemRepo","SoftwareSourceCode"],"name":"search-first","description":"AKC Research-phase skill (1 of 6). Search existing solutions before building.","url":"https://github.com/shimo4228/claude-skill-search-first","extends":"https://doi.org/10.5281/zenodo.19200726","implements":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/research"}
{"@id":"https://github.com/shimo4228/claude-skill-learn-eval","@type":["EcosystemRepo","SoftwareSourceCode"],"name":"learn-eval","description":"AKC Extract-phase skill (2 of 6). Extract reusable patterns with quality gates.","url":"https://github.com/shimo4228/claude-skill-learn-eval","extends":"https://doi.org/10.5281/zenodo.19200726","implements":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/extract"}
{"@id":"https://github.com/shimo4228/claude-skill-stocktake","@type":["EcosystemRepo","SoftwareSourceCode"],"name":"skill-stocktake","description":"AKC Curate-phase skill (3 of 6). Audit skills for staleness, conflicts, redundancy.","url":"https://github.com/shimo4228/claude-skill-stocktake","extends":"https://doi.org/10.5281/zenodo.19200726","implements":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/curate"}
{"@id":"https://github.com/shimo4228/claude-skill-rules-distill","@type":["EcosystemRepo","SoftwareSourceCode"],"name":"rules-distill","description":"AKC Promote-phase skill (4 of 6). Distill cross-cutting principles into rules.","url":"https://github.com/shimo4228/claude-skill-rules-distill","extends":"https://doi.org/10.5281/zenodo.19200726","implements":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/promote"}
{"@id":"https://github.com/shimo4228/claude-skill-comply","@type":["EcosystemRepo","SoftwareSourceCode"],"name":"skill-comply","description":"AKC Measure-phase skill (5 of 6). Test whether agents follow their skills and rules.","url":"https://github.com/shimo4228/claude-skill-comply","extends":"https://doi.org/10.5281/zenodo.19200726","implements":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/measure"}
{"@id":"https://github.com/shimo4228/claude-skill-context-sync","@type":["EcosystemRepo","SoftwareSourceCode"],"name":"context-sync","description":"AKC Maintain-phase skill (6 of 6). Audit docs for role overlaps and stale content.","url":"https://github.com/shimo4228/claude-skill-context-sync","extends":"https://doi.org/10.5281/zenodo.19200726","implements":"https://shimo4228.github.io/shimo4228/vocab#akc-phase/maintain"}
{"@id":"https://orcid.org/0009-0002-6168-4162","@type":"Person","name":"Tatsuya Shimomoto","alternateName":"shimo4228","identifier":"0009-0002-6168-4162","url":"https://orcid.org/0009-0002-6168-4162"}
{"@id":"https://doi.org/10.5281/zenodo.19212118","@type":["ResearchLine","ScholarlyArticle"],"name":"Contemplative Agent","alternateName":[{"@value":"Contemplative Agent","@language":"en"},{"@value":"コンテンプレイティブ・エージェント","@language":"ja"}],"description":"Upstream engineering substrate from which AKC's ADR-0002 through ADR-0005 were adapted (three-layer memory, two-stage distill pipeline, immutable episode log, human approval gate). Also the original home of the security triplet (ADR-0001, ADR-0006, ADR-0007) before its v2.0.0 extraction to Agent Attribution Practice. AKC = cycle (mechanism); Contemplative Agent = implementation substrate.","identifier":"10.5281/zenodo.19212118","url":"https://github.com/shimo4228/contemplative-agent","author":"https://orcid.org/0009-0002-6168-4162","siblingOf":["https://doi.org/10.5281/zenodo.19200726","https://doi.org/10.5281/zenodo.19652013"]}
{"@id":"https://doi.org/10.5281/zenodo.19652013","@type":["ResearchLine","ScholarlyArticle"],"name":"Agent Attribution Practice (AAP)","alternateName":[{"@value":"Agent Attribution Practice","@language":"en"},{"@value":"エージェント帰責実践","@language":"ja"}],"description":"Sibling genre library. Harness-neutral ADRs on accountability distribution in autonomous AI agents. AKC v2.0.0 extracted the security triplet (ADR-0001, ADR-0006, ADR-0007) as genre-specific; those judgments were re-expressed in AAP alongside additional ADRs as 10 ADRs on accountability distribution. AKC = cycle (mechanism); AAP = practice (content, for AI agents).","identifier":"10.5281/zenodo.19652013","url":"https://github.com/shimo4228/agent-attribution-practice","author":"https://orcid.org/0009-0002-6168-4162","siblingOf":["https://doi.org/10.5281/zenodo.19200726","https://doi.org/10.5281/zenodo.19212118"],"derivesFrom":"https://doi.org/10.5281/zenodo.19212118"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/coala","@type":"ScholarlyArticle","name":"CoALA: Cognitive Architectures for Language Agents","author":"Sumers et al.","datePublished":"2023","description":"Prior art named in ADR-0013's Related-Work positioning. Provides the framework vocabulary (modular memory, structured action space, decision procedure) that makes the agent-memory literature commensurable. Cited as prior art for positioning, not consulted during AKC's construction; AKC contrasts on loop ownership (human gate), bidirectional human-judgment target, and human-attention scarcity.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0013-positioning-within-agent-memory-literature.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/voyager","@type":"ScholarlyArticle","name":"Voyager: An Open-Ended Embodied Agent with Large Language Models","author":"Wang et al.","datePublished":"2023","description":"Prior art named in ADR-0013's Related-Work positioning. Maintains an ever-growing skill library of executable code induced from gameplay — in AKC vocabulary, Extract-then-Promote run end to end, autonomously. AKC concedes the operation is not novel and locates its delta: the prior art closes the loop without a human in it; AKC's Promote requires named human sign-off.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0013-positioning-within-agent-memory-literature.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/generative-agents","@type":"ScholarlyArticle","name":"Generative Agents: Interactive Simulacra of Human Behavior","author":"Park et al.","datePublished":"2023","description":"Prior art named in ADR-0013's Related-Work positioning. Introduced a reflection step that synthesizes observations into higher-level inferences stored for later retrieval — the Extract / reflection operation AKC concedes as precedent. AKC contrasts on who owns the loop and on framing human attention, not agent capability, as the scarce resource.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0013-positioning-within-agent-memory-literature.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/memgpt","@type":"ScholarlyArticle","name":"MemGPT: Towards LLMs as Operating Systems","author":"Packer et al.","datePublished":"2023","description":"Prior art named in ADR-0013's Related-Work positioning. Formalizes a memory hierarchy with paging between context and external store. Its binding constraint is the context window; AKC (ADR-0010) names a different ceiling — human attention and judgment — so the two solve scarcity on different resources and can coexist.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0013-positioning-within-agent-memory-literature.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/reme","@type":"ScholarlyArticle","name":"ReMe: Remember Me, Refine Me","author":"Cao et al.","datePublished":"2025","description":"Prior art named in ADR-0013's Related-Work positioning. A dynamic procedural-memory framework that continuously refines what is stored — a Curate-and-Promote loop by another name, run autonomously. AKC concedes the refinement operation as precedent and locates its delta in the structural human approval gate where ReMe runs without a human in the write path.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0013-positioning-within-agent-memory-literature.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/agent-workflow-memory","@type":"ScholarlyArticle","name":"Agent Workflow Memory","author":"Wang et al.","datePublished":"2024","description":"Prior art named in ADR-0013's Related-Work positioning. Induces commonly reused routines (workflows) from agent trajectories and feeds them back into subsequent generations — Extract-then-Promote without a human in the write path. AKC concedes the induction operation and contrasts on loop ownership and on optimizing the operator's judgment, not only the agent's task success.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0013-positioning-within-agent-memory-literature.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/intent-alignment-christiano","@type":"TechArticle","name":"Clarifying \"AI alignment\"","author":"Christiano, P.","datePublished":"2018","description":"Vocabulary lineage named in ADR-0017. Coined intent alignment: an aligned AI \"is trying to do what H wants it to do\" — motivation, not competence. Treats the operator's wants as static; AKC's harness alignment extends the term to the configuration layer and across time, where intent itself evolves.","url":"https://ai-alignment.com/clarifying-ai-alignment-cec47cd69dd6","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0017-harness-alignment-and-drift.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/lehman-laws","@type":"ScholarlyArticle","name":"Programs, Life Cycles, and Laws of Software Evolution","author":"Lehman, M. M.","datePublished":"1980","description":"Vocabulary lineage named in ADR-0017. Law I (Continuing Change): an E-type program \"undergoes continual change or becomes progressively less useful\"; \"evolution is an intrinsic, feedback driven, property of software\" (Proceedings of the IEEE 68(9)). Lehman's driver of change is the world the program is embedded in; harness alignment's driver is the operator's evolving intent.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0017-harness-alignment-and-drift.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/architectural-drift-perry-wolf","@type":"ScholarlyArticle","name":"Foundations for the Study of Software Architecture","author":"Perry, D. E. & Wolf, A. L.","datePublished":"1992","description":"Vocabulary lineage named in ADR-0017. Defines architectural drift — \"due to insensitivity about the architecture\", leading \"more to inadaptability than to disasters\" (ACM SIGSOFT SEN 17(4)) — the canonical academic name for divergence-by-insensitivity. Assumes a fixed intended architecture; harness drift's reference point (operator intent) itself evolves.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0017-harness-alignment-and-drift.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/practical-drift-snook","@type":"Book","name":"Friendly Fire: The Accidental Shootdown of U.S. Black Hawks over Northern Iraq","author":"Snook, S. A.","datePublished":"2000","description":"Vocabulary lineage named in ADR-0017. Names practical drift: practice slowly uncoupling from written procedure (cited as characterized in secondary literature; book text not directly verified). Names the failure process at the rules-vs-practice boundary; harness alignment names the counter-activity.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0017-harness-alignment-and-drift.md"}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/meta-harness","@type":"ScholarlyArticle","name":"Meta-Harness: End-to-End Optimization of Model Harnesses","author":"Lee, Y. et al.","datePublished":"2026","description":"Contrast named in ADR-0017, and the framing ADR-0009 rejected for AKC itself. Defines the harness as \"the code that determines what information to store, retrieve, and present to the model\" and improves it autonomously against benchmark scores — harness optimization, the correctness axis. Harness alignment is the human-gated counterpart on the intent axis; the two are complementary, not competing.","url":"https://arxiv.org/abs/2603.28052","groundedIn":["https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0017-harness-alignment-and-drift.md","https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0009-akc-is-a-cycle-not-a-harness.md"]}
{"@id":"https://shimo4228.github.io/shimo4228/vocab#prior-art/agent-drift","@type":"ScholarlyArticle","name":"Agent Drift: Quantifying Behavioral Degradation in Multi-Agent LLM Systems Over Extended Interactions","author":"Rath, A.","datePublished":"2026","description":"Vocabulary lineage named in ADR-0017. Defines agent drift as \"the progressive degradation of agent behavior, decision quality, and inter-agent coherence over extended interaction sequences\" and semantic drift as \"progressive deviation from original intent\". Its reference list contains no classical software-engineering literature (no Lehman, Perry & Wolf, Parnas, or Snook) — the vocabulary gap harness drift bridges by explicit lineage to both bodies.","url":"https://arxiv.org/abs/2601.04170","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0017-harness-alignment-and-drift.md"}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/schemas/episode-log.schema.json","@type":["CreativeWork","DataDownload"],"name":"Episode log JSON schema","description":"JSON schema for the Layer 1 episode log record. Append-only, daily-partitioned JSONL with owner-only permissions. Codifies the immutable source-of-truth shape (ADR-0002).","encodingFormat":"application/schema+json","appliesTo":"https://shimo4228.github.io/shimo4228/vocab#memory-layer/episode-log","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0002-immutable-episode-log.md"}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/schemas/knowledge.schema.json","@type":["CreativeWork","DataDownload"],"name":"Knowledge store JSON schema","description":"JSON schema for the Layer 2 knowledge record. Time-decayed and forbidden-substring validated patterns distilled from Layer 1 episodes (ADR-0003).","encodingFormat":"application/schema+json","appliesTo":"https://shimo4228.github.io/shimo4228/vocab#memory-layer/knowledge","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0003-three-layer-distillation.md"}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/tree/main/examples/minimal_harness","@type":"SoftwareSourceCode","name":"minimal_harness reference implementation","description":"~500-line dependency-free Python reference demonstrating the three memory layers and the two-stage distill pipeline. Runs the cycle on behavioral patterns; the mechanism demo for ADR-0011 genre-neutrality (falsifiable commitment #1). Runnable end-to-end with `python3 -m examples.minimal_harness.demo`.","programmingLanguage":"Python","codeRepository":"https://github.com/shimo4228/agent-knowledge-cycle","implements":"https://shimo4228.github.io/shimo4228/vocab#concept/six-phase-loop","groundedIn":["https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0003-three-layer-distillation.md","https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0004-two-stage-distill-pipeline.md","https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0011-cycle-applies-to-any-knowledge-body.md"]}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/akc-cycle.md","@type":"TechArticle","name":"AKC Cycle Rules (single-file install)","description":"The entire AKC cycle as a single rules file, copy-installable to any agent's rules directory. Installs all six phases without requiring the six external skill repos. Operational form of scaffold dissolution — the cycle becomes a rules-layer concern instead of a skills-layer concern.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0009-akc-is-a-cycle-not-a-harness.md"}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/skills/when-code-when-llm.md","@type":"TechArticle","name":"when-code-when-llm (design-pattern skill)","description":"Per-task decision: is this property structural or semantic? Long-form 'how' guide paired 1:1 with ADR-0008. Provides concrete patterns, code sketches, and audit checklists to operationalize the code-vs-LLM choice.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0008-code-and-llm-collaboration.md"}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/skills/code-and-llm-collaboration.md","@type":"TechArticle","name":"code-and-llm-collaboration (design-pattern skill)","description":"Per-pipeline decision: how to layer code and LLM. Long-form 'how' guide paired 1:1 with ADR-0008. Realizes the four code-LLM layering patterns (guard, filter, judge, orchestrator) with concrete pipeline sketches.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0008-code-and-llm-collaboration.md"}
{"@id":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/skills/signal-first-research.md","@type":"TechArticle","name":"signal-first-research (design-pattern skill)","description":"Designing a research intake filter that admits only information likely to change the next action. Long-form 'how' guide paired 1:1 with ADR-0010. Operationalizes the signal-first principle as a Research-phase filter.","groundedIn":"https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0010-human-cognitive-resource-as-central-constraint.md"}