Glossary

Definitions for the terms, concepts, and infrastructure components used throughout the Cognitive Substrate series. Click any term for a full description.

A

Abstraction engine: A background worker that clusters experience events into progressively higher-order concepts using embedding-based similarity. Produces a five-level compression ladder: experience - pattern - concept - principle - worldview. Vendor-agnostic at every level.
Activity trace: A structured record emitted by each agent during a cognitive loop iteration. Captures the agent's role, inputs, outputs, confidence estimate, and timing. Stored in the agent activity index and used for meta-cognitive attribution and performance analysis.
Affect state: A persistent vector representing the system's current emotional context across dopamine-like (reward sensitivity), norepinephrine-like (urgency and arousal), and serotonin-like (stability) channels. Affect state modulates attention salience and reinforcement weighting in real time.
Agent: A specialized cognitive worker that produces a proposal, evaluation, prediction, or retrieval result under a shared runtime contract. The cognitive loop coordinates a planner agent, executor agent, critic agent, memory agent, and meta-cognition agent as standard roles.
Agent context: The input bundle provided to an agent on each invocation, including a session snapshot, retrieved memories, active goals, policy state, and the current input payload.
Arbitration: The process that compares proposals from multiple agents, scores candidates across four dimensions (coherence, predicted reward, memory alignment, risk), and selects a final decision for execution. Memory alignment is capped at five memories to prevent quantity from substituting for quality.
Architecture mutation: A structural change to the cognitive system's components, roles, strategies, or subsystem relationships proposed during open-ended evolution mode. All mutations pass constitutional checks before taking effect and are introduced at low weight, with reinforcement determining whether they improve performance.
Associative memory layer: The OpenSearch-backed retrieval layer that stores searchable metadata, vectors, semantic abstractions, memory links, and retrieval feedback. The primary interface for active recall during the cognitive loop.
Attention engine: The subsystem that allocates cognitive resources across competing signals based on a six-dimensional salience score: importance, relevance, urgency, novelty, risk, and goal relevance. Manages a working-memory budget of five primary and ten background items.
Audit stream: An immutable Kafka event stream that records state-changing events - memory updates, policy updates, self-modification outcomes - for compliance, debugging, and replay.
Avro: A binary serialization format used with the Schema Registry to encode Kafka messages compactly and with schema enforcement. Avro messages include a schema ID rather than the full schema, keeping message payloads small while still allowing consumers to retrieve and validate the full structure.

B

Backpressure: A flow-control condition in which a downstream consumer cannot process messages as fast as an upstream producer publishes them, causing queue depth to grow. Detected as the BACKPRESSURE_ACCUMULATION operational primitive. Unaddressed backpressure leads to lag growth, latency expansion, and eventually broker or consumer saturation.
BM25: A keyword-based relevance ranking algorithm used in the memory gateway's hybrid retrieval pipeline. BM25 scores documents by term frequency and inverse document frequency, complementing k-NN vector similarity to ensure lexically distinctive terms are retrievable even when embedding similarity is low.
Branching cognition: A runtime model in which divergent lines of reasoning, interruptions, and competing contextual continuations are represented as traceable branches with lifecycle state and explicit merge or abandon decisions.
Budget engine: The subsystem that allocates cognitive resources (compute, retrieval depth, reasoning complexity) to operations based on utility threshold gating. Controls fast-mode vs. slow-mode selection and enforces graceful degradation under resource exhaustion.

C

Calibration error: The difference between an agent's stated confidence and its actual accuracy on that operation. Computed as the absolute deviation: |confidence - accuracy|. High mean calibration error across operations indicates the system is systematically overconfident or underconfident, which misleads the arbitration scoring that uses confidence as a 30% weight.
Capability search: The open-ended evolution process that searches for structural improvements to the cognitive system when policy learning has converged and performance remains insufficient. Triggered by repeated failure, curiosity pressure, or persistent contradictions. Guided by the curiosity engine and constrained by constitutional checks.
Cassimatis's cognitive substrate hypothesis: The proposal by Nicholas Cassimatis that a small set of computational mechanisms - substrate-level operations such as focus of attention, unification, and constraint propagation - can explain the full range of human cognitive capabilities when combined and applied recursively. The hypothesis influenced the name and conceptual framing of the Cognitive Substrate project.
Causal model: A structural model inferred from co-occurrence in experience history that represents directional relationships between operational events. Supports counterfactual reasoning and intervention evaluation.
ClickHouse: The append-only analytical datastore used as the temporal telemetry intelligence layer. Stores raw metrics, logs, traces, cognitive events, and pattern outcomes. Optimized for aggregation-heavy time-series queries and intended to operate at 1,000+ service scale.
Cognitive bus: The Kafka event fabric treated as the temporal coordination layer for asynchronous cognition. All cognitive subsystems communicate via typed Kafka topics rather than direct in-process calls.
Cognitive loop: The six-step per-event cycle that constitutes active cognition: perceive - retrieve - reason - act - evaluate - consolidate. The orchestrator coordinates this loop; each step is observable and auditable.
Cognitive observability: Observability focused on cognition as a distributed process. Covers memory retrieval decisions, arbitration margins, policy drift vectors, consolidation results, and self-modification proposals using OpenTelemetry under the cog.* namespace.
Cognitive session: A runtime-scoped state container that persists trace context, working memory references, policy snapshot, participating agents, and arbitration state across the duration of an interaction.
Coherence score: One of four dimensions scored by the arbitration process when evaluating agent proposals. Measures whether the proposal is internally consistent and compatible with the current task, active goals, and policy state. Incoherent proposals can be rejected by the critic before reaching arbitration.
Confabulation: The failure mode in which synthetic or imagined events become indistinguishable from observed facts, corrupting the world model and memory substrate with unverified content. The dream engine prevents confabulation by tagging all synthetic events and blocking them from the episodic truth layer.
Consolidation: A batch-oriented pipeline stage that replays episodic events to produce semantic abstractions, reinforcement updates, and decay decisions. Runs as a background worker on a configurable interval; the interval must be short enough to prevent catastrophic convergence of trusted memories.
Constitutional engine: The invariant policy layer that checks proposed actions, policy updates, and self-modification proposals before they take effect. Triggers quarantine when identity drift exceeds a configured threshold or when two independent reward-corruption signatures are detected simultaneously.
Consumer group: A named group of Kafka consumers that jointly consume a topic, with each partition assigned to exactly one consumer in the group at a time. Consumer groups enable parallel processing and horizontal scaling of pipeline workers. Group lag (the difference between the latest offset and the committed offset) is the primary autoscaling signal for KEDA.
Consumer lag: The number of unprocessed messages between a consumer group's committed offset and the topic's latest offset. Rising lag indicates a consumer cannot keep pace with the producer and is the primary autoscaling signal for KEDA-managed workers.
Contradiction risk: A scoring channel in the reinforcement engine that reduces reward when an experience event conflicts with existing high-trust memories. Prevents contradictory or low-quality memories from accumulating strength through Hebbian compounding despite disagreeing with the established knowledge base.
Counterfactual: A hypothetical query about what would have happened under different conditions, evaluated against the causal model. Used by the grounding engine and world model to assess the impact of interventions and distinguish genuine causal relationships from mere correlations.
Cross-encoder: A reranking model that scores query-document pairs jointly rather than encoding them independently. Used as an optional second-pass step in the memory gateway to improve precision after an initial BM25 and k-NN retrieval pass returns a candidate set.
Curiosity engine: The subsystem that proposes bounded experiments to reduce uncertainty in high-value areas. Curiosity score is a weighted combination of information gain (40%), novelty (25%), uncertainty (25%), and visit count (10%). All experiments pass constitutional and budget checks before execution.

D

Debate trace: A record of the structured deliberation between the critic and planner agents during the evaluation phase of the cognitive loop. Captures disagreements, risk flags, confidence margins, and the final arbitration decision. Used by the reflection loop to attribute reasoning failures to specific deliberation steps.
Decay: A forgetting mechanism that reduces the retrieval priority and survival probability of low-utility memories over time. Decay is multiplicative, which means Hebbian compounding gains erode without periodic re-consolidation.
Developmental stage: One of five capability phases the system advances through based on demonstrated readiness: seed, novice, apprentice, integrative, and open-ended. Higher-capability operations are gated on evidence of proficiency at the preceding stage, not elapsed time.
Diskless topic: A Kafka topic backed by object storage rather than broker-local disks. Brokers become stateless routing nodes. Suitable for high-volume append-only streams where object-storage latency is acceptable and long retention is required at low cost.
Dopamine-like signal: The reward-sensitivity channel of the affect state vector, modeled loosely on dopaminergic signaling. High values increase attention salience for novel and high-reward signals and amplify reinforcement weight for positive outcomes. Distinct from the serotonin-like stability channel and norepinephrine-like urgency channel.
Dream engine: The subsystem that synthesizes hypothetical scenarios from memories, goals, contradictions, and curiosity targets to test world-model predictions offline. Synthetic events are tagged as synthetic to prevent confabulation - imagined scenarios becoming indistinguishable from facts.

E

Embedding: A vector representation of text or structured input used for similarity search and retrieval in the memory substrate. The system supports multiple embedding profiles simultaneously: quality (higher accuracy), efficient (lower latency), and hybrid (blend of both).
Emergence record: A structured event documenting an accepted or rejected architecture mutation, the conditions that triggered capability search, and the observed outcomes. Accumulated emergence records allow the meta-cognition engine to distinguish beneficial structural evolution from destabilizing drift.
Enrichment: A processing step that adds embeddings and metadata - tags, importance scoring, and initial reward values - to raw experience events before they are indexed.
Episodic truth layer: The immutable object-storage archive for full-fidelity experience payloads and raw reasoning traces. Indexed metadata in OpenSearch points to payloads here via object_storage_key.
Experience event: The atomic unit of cognition: a structured record binding an input, the internal state at the time, the action taken, the outcome, and an initial evaluation score. The evaluation score seeds the reinforcement pipeline. In code: ExperienceEvent.
Experience ingestion: The pipeline stage that accepts incoming experience events and publishes them to the event bus for downstream enrichment, indexing, reasoning, and evaluation.
Exploration factor: A policy engine weight that controls the balance between exploiting known successful strategies and exploring novel approaches. Adjusted by reinforcement based on recent reward history. Higher values increase the probability of selecting less-proven strategies during arbitration.

F

Fast mode: A budget engine operating mode that reduces retrieval depth, reasoning complexity, and the number of agents dispatched to meet a latency target. Activated when computational resources are constrained or when the current operation's utility score falls below the threshold warranting full slow-mode processing.
Forgetting system: The set of mechanisms that suppress, compress, retire, and prune memories to prevent unbounded growth and catastrophic convergence. Operates via severity-stratified suppression, re-consolidation windows, compression into semantic abstractions, and graph pruning.

G

Goal relevance: A salience dimension and reinforcement channel that measures how directly an experience event or memory relates to currently active goals. High goal relevance increases retrieval priority and reinforcement weight, biasing the system toward evidence that bears on its current objectives.
Goal system: The representation of long-horizon objectives, organized into a hierarchy of five timescales. Goals feed progress signals to reinforcement independent of immediate outcomes and bias retrieval toward relevant memories.
Grounding engine: The subsystem that compares external telemetry against world-model predictions to compute prediction error, correct model confidence estimates, and trigger active inference probes in high-uncertainty areas.

H

Hebbian compounding: The mechanism by which memories accumulate reinforcement priority across repeated retrieval and positive reinforcement. Implemented as a logarithmic count bonus: finalRp += countBonus x log2(1 + count) x reinforcement. A quality gate multiplies the bonus by reinforcement signal strength so contradictory memories cannot accumulate by volume alone.
HNSW: Hierarchical Navigable Small World - a graph-based approximate nearest neighbor index algorithm used in OpenSearch for fast k-NN vector search. HNSW indexes enable millisecond retrieval across millions of embeddings by navigating a multi-layer proximity graph rather than scanning all vectors.
Horizontal scaling: A scaling property in which agent workers and pipeline components scale by adding parallel replicas rather than larger machines, coordinated via Kafka partitions and consumer groups.

I

Identity drift: The temporal evolution of identity state in response to reinforcement signals, experience history, and policy updates. Measured as RMS distance from the established identity vector across six behavioral dimensions. Quarantine triggers when drift exceeds maxIdentityDrift.
Identity state: A persistent, slowly-evolving vector representation of behavioral tendencies - curiosity, caution, verbosity, tool dependence, adaptability, and stability. Distinct from policy: policy weights adapt rapidly to local reinforcement; identity drifts slowly from long-term behavioral patterns.
Importance score: A scalar used to prioritize retention, indexing, consolidation, and retrieval of experience events and semantic memories.
Incremental consolidation: A consolidation strategy that processes only newly arrived or recently modified experience events rather than scanning the entire memory corpus on each run. Required at scale where full-corpus consolidation is computationally prohibitive. Uses watermarks or change logs to track processing boundaries.
Introspection: The meta-cognitive process of inspecting the system's own reasoning traces, confidence estimates, and calibration history to identify failure modes and propose improvements. Treated as a resource with a cost that must be economically justified, not an unlimited entitlement.

K

k-NN: k-Nearest Neighbor search - the vector similarity retrieval method used in the memory gateway to find memories whose embeddings are closest to the query embedding. Combined with BM25 keyword search in the hybrid retrieval pipeline. Accelerated by HNSW indexing in OpenSearch.
Kafka topic: A named, ordered, partitioned log of events in the Kafka messaging system. Producers publish to topics; consumer groups consume from them. Topics are the fundamental unit of data flow in the cognitive bus, organized into tiered namespaces (telemetry.*, cognition.*, memory.*, selfmod.*).
KEDA: Kubernetes Event-Driven Autoscaler. Used to scale cognitive workers based on Kafka lag and other stream signals.

M

Memory alignment: One of four arbitration scoring dimensions. Measures how many retrieved memories support a given agent proposal, capped at five to prevent quantity from outweighing quality.
Memory critique: An event that flags a specific memory as inconsistent, outdated, or contradicted by more recent evidence. Generated by the critic agent, the reflection loop, or the grounding engine. Feeds the forgetting system's suppression and re-consolidation decisions.
Memory gateway: The component that performs retrieval from the memory substrate and provides a normalized interface to the cognitive loop and agents. Supports hybrid retrieval (BM25 + k-NN) with optional cross-encoder reranking.
Memory links: Graph relationships between memories that support multi-hop retrieval and structural navigation. Link types include supports, contradicts, extends, and associates. Stored in the memory_links OpenSearch index.
Memory substrate: The full persistent storage and retrieval infrastructure: OpenSearch for associative retrieval, object storage for immutable full-fidelity payloads, and ClickHouse for telemetry analytics.
Meta-cognition engine: The subsystem that monitors reasoning quality, measures calibration error, attributes failures by operation type, and emits self-modification proposals. Operates under a budget constraint: introspection has a cost and must be economically justified like any other cognitive operation.

N

Narrative engine: The subsystem that synthesizes autobiographical threads from identity state and reinforcement history, scores them for coherence, and projects a future-self trajectory. Narrative coherence is a measure of how far current behavior has moved from the established self-model.
Norepinephrine-like signal: The urgency and arousal channel of the affect state vector. High values increase the salience weight of time-sensitive and high-risk signals and compress deliberation budget toward faster decisions. Elevated during high-stakes or time-constrained operational contexts.
Novelty: A retrieval weighting signal computed as 1 - decay^(t - lastSeen(id)). Memories not recently accessed during the current session score as novel, promoting retrieval diversity and preventing attentional fixation on the same high-importance memories.

O

Object storage: A cloud-native blob storage service (S3 or equivalent) used as the episodic truth layer for full-fidelity experience payloads and as the backing store for diskless Kafka topics. Optimized for high durability and low cost at the expense of retrieval latency compared to local disk or in-memory stores.
Offset: A monotonically increasing integer that uniquely identifies a message's position within a Kafka partition. Consumer groups commit their current offset to track progress. Resetting an offset to an earlier position enables replay of historical events without re-publishing them.
Open-ended evolution: The operating mode in which the cognitive system is authorized to propose and trial structural capability changes beyond policy adaptation. Activated only when the developmental engine confirms the open-ended capability phase and when policy learning has converged without resolving persistent failures.
OpenSearch: The associative memory store used for keyword and vector retrieval across experience events, semantic memories, policy state, agent activity, world-model predictions, goals, and identity state.
OpenTelemetry: The observability framework used to instrument cognition as distributed traces, metrics, and logs. Project-specific semantic conventions live under the cog.* namespace.
Operational primitive: A system-agnostic unit of distributed system behavior that forms the basic vocabulary of the pattern library. Examples: BACKPRESSURE_ACCUMULATION, QUEUE_GROWTH, TAIL_LATENCY_EXPANSION. Primitives describe behavioral dynamics rather than vendor metrics, making learned patterns portable across infrastructure environments.

P

Partition: A subdivision of a Kafka topic that enables parallel processing. Each partition is an ordered, independent log consumed by at most one consumer in a consumer group at a time. Partition count determines the maximum achievable parallelism for a topic's consumer group.
Pattern library: The OpenSearch-backed store of system-agnostic operational patterns. Each pattern defines a signature (co-occurring primitives), precursors (early-warning primitives), an outcome description, and ordered interventions. Confidence is updated by the reinforcement feedback worker based on observed recommendation outcomes.
Planner agent: The multi-agent role responsible for generating a proposed strategy given the current context. Receives episodic memories, active goals, and policy state. Does not evaluate whether the plan is good - that is the critic's role. Its proposal enters arbitration alongside the executor's and critic's outputs.
Policy alignment: A reinforcement channel measuring how consistent an experience event is with the current policy state. High alignment contributes positively to the reinforcement score; low alignment reduces it. Weighted at 16% in the reinforcement scoring function.
Policy convergence: The state in which the policy engine's adaptive weights have stabilized and further reinforcement produces diminishing changes. One of the trigger conditions for open-ended evolution: if policy convergence occurs while the system still experiences persistent failure, it indicates the current architecture's capability ceiling has been reached.
Policy drift: The time-varying change in policy state driven by reward signals. Per-step drift is bounded by MAX_ABSOLUTE_DRIFT = 0.08 to prevent rapid destabilization. Total accumulated drift over a session is auditable.
Policy engine: The subsystem that maintains adaptive behavioral weights across dimensions including retrieval weighting, exploration preference, risk tolerance, and arbitration bias. Updates are reward-driven and bounded.
PostgreSQL: The relational coordination store used for state that benefits from referential constraints: agent configuration, identity anchors, session state, and policy version history.
Prediction accuracy: A scalar measuring how closely a world-model forecast matched an observed outcome: 1 - error / scale where scale = max(1, |expected|). A latency prediction of 50ms against an observed 1200ms produces an accuracy score of 0.000, which feeds directly into world-model correction.

Q

Quarantine: An isolation state applied to memories, policy updates, or self-modification proposals that triggered a constitutional violation or identity drift threshold breach. Quarantined items are excluded from active retrieval and execution but preserved in the audit stream for review and potential resubmission after stabilization.

R

Reinforcement: The process that computes reward signals from operation outcomes and propagates reward-driven updates into retrieval priority, policy state, and identity state. Uses Hebbian compounding with a quality gate.
Reinforcement feedback worker: The operational worker that tracks pending and completed recommendations, computes exponential moving average confidence updates per pattern, and writes updated confidence values to both the serving index and the audit trail.
Replay: The reprocessing of historical Kafka events by resetting a consumer group's offset to an earlier position. Used for re-indexing after schema changes, re-running consolidation with updated logic, and debugging pipeline failures by reproducing the exact event sequence that caused them.
Retrieval feedback: A recorded signal evaluating the utility of a specific retrieval result. Used to improve future retrieval weighting and reinforcement priority. Stored in the retrieval_feedback OpenSearch index.
Retrieval priority: A scalar score maintained per memory that determines how prominently it surfaces in k-NN and hybrid retrieval results. Computed as a weighted combination of reinforcement signal, usage frequency, and goal relevance. Increased by Hebbian compounding on positive reinforcement; reduced by decay over time.
Reward corruption: A failure mode in which the reinforcement signal itself is manipulated or miscalibrated, causing the system to optimize for a proxy rather than the intended outcome. The constitutional engine requires two independent corruption signatures before triggering quarantine, preventing false positives from a single noisy signal.

S

Salience: The composite signal that determines which items receive cognitive resources. Combines six dimensions: importance, relevance, urgency, novelty, risk, and goal relevance. Affect state modulates the weighting of individual dimensions based on environmental context.
Schema evolution: The process of changing a message schema over time while maintaining compatibility with existing producers and consumers. Governed by the Schema Registry's compatibility rules: backward compatibility allows old consumers to read new messages; forward compatibility allows new consumers to read old messages; full compatibility satisfies both.
Schema Registry: A central repository that stores the formal schema definitions for every message type flowing through the Kafka pipeline. Producers validate messages against registered schemas before publishing; consumers retrieve schemas to deserialize received messages. Enforces compatibility rules to prevent breaking changes from reaching production.
Self-modification: A controlled proposal, validation, and application of architecture or policy mutations produced by meta-cognitive mechanisms. All proposals pass constitutional checks before taking effect.
Semantic memory: A consolidated abstraction derived from episodic replay: a distilled representation of patterns, principles, or recurring observations extracted from raw experience events. Stored in the memory_semantic OpenSearch index.
Semantic similarity: A measure of how closely related two pieces of text are in meaning, computed as the distance (typically cosine similarity) between their embeddings in vector space. The primary signal for k-NN retrieval in the memory gateway. High semantic similarity does not require lexical overlap.
Serotonin-like signal: The stability channel of the affect state vector. High values bias the system toward consistency, caution, and established patterns. Low values increase receptivity to novel or disruptive signals. Counterbalances the dopamine-like channel's reward-seeking tendency.
Session manager: The runtime component responsible for the lifecycle of cognitive sessions: creation, state hydration, persistence between turns, and cleanup.
Slow mode: A budget engine operating mode that applies full retrieval depth, all agent roles, cross-encoder reranking, and maximum reasoning complexity. Used when the utility of a decision justifies the computational cost. The default for high-stakes or high-uncertainty operations.
Social engine: The subsystem that maintains persistent models of users and peers, infers intent and belief states, and tracks trust and deception risk through separate update mechanisms.
System mapping: An adapter definition that binds vendor-specific metric names to the operational primitive vocabulary. When a new infrastructure environment is onboarded, only the mapping changes; the pattern library and detection logic remain unchanged.

T

Telemetry topic tier: The telemetry.* Kafka namespace. Carries raw infrastructure telemetry: telemetry.metrics.raw, telemetry.logs.raw, telemetry.traces.raw, and the normalized derivative telemetry.events.normalized. All topics use Diskless storage because scrape intervals are 15 seconds or longer.
Temporal engine: The subsystem that represents time as a cognitive dimension: planning horizons, urgency gradients, and episodic sequencing. Allocates reasoning depth based on temporal distance and deadline pressure.
Tiered storage: A storage architecture that assigns memories to hot, warm, or cold tiers based on recency, trust score, and retrieval frequency. Hot memories live in fully-indexed OpenSearch with HNSW acceleration; warm memories use standard indexing; cold memories are archived in object storage. Movement between tiers is automated.
Transfer layer: The combination of the system mapping DSL and the pattern library that enables learned operational intelligence to be applied to a new infrastructure environment without retraining. When pointed at a new environment, only a SystemMapping is required; the pattern worker immediately begins matching incoming primitives against the existing pattern library.
Trust score: A per-memory scalar that evolves over time based on reinforcement outcomes. Rises when the memory contributes to correct, helpful results; falls when it leads to corrections or poor outcomes. The primary quality signal used by tiered storage assignment, retrieval priority weighting, and the forgetting system's suppression decisions.

U

Usage frequency: A retrieval-priority input tracking how often a memory has been retrieved in recent sessions. Combined with reinforcement signal and goal relevance to compute retrieval priority. High usage frequency partially suppresses the novelty bonus to prevent frequently-retrieved memories from accumulating an additional novelty credit.

V

Vector search: Retrieval based on embedding similarity rather than keyword matching. Finds memories that are semantically related to a query even when they share no exact terms. Implemented in OpenSearch using HNSW indexes for approximate nearest neighbor lookup. The k-NN component of the memory gateway's hybrid retrieval pipeline.

W

Watchdog agent: A meta-cognitive monitoring agent that tracks a specific signal (calibration error, identity drift, policy drift magnitude, retrieval quality) and emits alerts or quarantine recommendations when the signal exceeds a configured threshold. Part of the Stage 20 meta-cognition control layer.
Working memory: A short-lived store for active context during a cognitive session: recent memories, active plans, and unresolved contradictions. Managed by the attention engine's budget constraints. Implemented in Redis or in-process cache.
World model: A predictive simulation subsystem that evaluates candidate actions prior to execution and emits risk and confidence estimates. Accuracy improves through prediction-error feedback from the grounding engine.