Skip to content

Canonical Vocabulary


Justification. Synixolis draws from multiple bodies of work that use overlapping terms with different meanings:

  • CDR theory uses “Level 1/2/3” for memory partitions, “Compress/Divide/Renew” as theorem names, and information-theoretic vocabulary (rate-distortion, entropy rate, sufficient statistic).
  • AAP v0.4.3 uses “Layer 0/1/2/3” for context model levels, “division” for spawning, and BEAM/OTP process vocabulary.
  • The ring architecture uses “Ring-0/1/2/3” for tiers and biological metaphors (senescence, pollution, cell division, hot/cold manifold).
  • Distributed systems literature uses “promotion” for upward data movement, “consensus” for agreement protocols, and “fencing” for fault containment.
  • Cognitive science uses “working memory,” “episodic memory,” and “semantic memory” for hierarchy levels that map to but do not coincide with the tier boundaries.

Without a canonical vocabulary, the spec is ambiguous. A reader encountering “promotion” cannot know whether it means upward tier movement (GC/HSM standard), spawning a handler agent (legacy kernel usage), or convention compilation (the optimization ladder). This is the same problem as namespace collision in distributed systems — resolve it once, reference the resolution everywhere.

Operating timescale. The vocabulary is static. It changes only when the spec is revised. Terms defined here are the schema, not the data — they are the IDL (Interface Definition Language) for the Synixolis protocol. All other sections of the spec, all conforming implementations, and all inter-node messages use these terms with these meanings.

How to use this section. This is a reference section, organized for lookup by domain. Each term carries a canonical name, definition, collision notes where terminology conflicts exist, and cross-references to the spec sections that depend on it. The collision resolution log (Section 2.9) documents every case where Synixolis terminology diverges from common usage and why.


The system organizes memory into four tiers with physics-appropriate lifecycle at each level. Tier naming follows CPU cache / HSM conventions (fastest to slowest, smallest to largest).

Canonical NameTimescaleDefinitionCognitive AnalogUsed In
Tier-0: Executionms to minAgent frame (stack + heap). KV cache, context window. The AAP protocol layer.Working memory (Baddeley)S3 (Node Contract), S8 (Reference Impl)
Tier-1: Sessionmin to hrsAgent state: messages, scope, heap across turns. The working set for a single agent lifetime.Short-term / episodic bufferS3, S5 (Colony), S8
Tier-2: Experiencehrs to wksArchive entries. Batch pipeline output. Episodic consolidation.Episodic memory (Tulving)S3, S4 (Optimization), S8
Tier-3: Identitywks to permanentCharter + receptor of long-lived agent lines. Core beliefs, stable preferences.Semantic memory (Tulving)S3, S5, S7 (Self-Modification), S8

Collision notes:

SourceOld TermResolution
CDR position paperLevel 1 / Level 2 / Level 3Remapped to Tier-1 / Tier-2 / Tier-3. CDR had no explicit Level 0; Tier-0 was implicit as “within Level 1.”
Ring architectureRing-0 / Ring-1 / Ring-2 / Ring-3Direct remap: Ring-N becomes Tier-N.
AAP v0.4.3Layer 0 (Task/Frame) / Layer 1 (Agent) / Layer 2 (Species)Remapped to Tier-0 / Tier-1 / Tier-2. AAP “Species” = Tier-2, not Tier-3.

Why “Tier” over “Ring”: “Ring” in OS convention refers to privilege levels (Ring 0 = kernel mode, Ring 3 = user mode), implying an access-control hierarchy. Misleading for memory tiers. “Tier” is the standard term in Hierarchical Storage Management and tiered caching, which is the actual analogy.

Why 0-indexed: CS convention (L1/L2/L3 caches, Tier-0/1/2 storage). Preserves the existing numbering from the ring architecture while fixing the name.


Standard terms from cache/HSM literature (Hennessy & Patterson, HSM specs).

Canonical NameDirectionDefinitionUsed In
ConsolidationTier-N to Tier-(N+1)Compressing and moving state from a faster tier to a slower tier. Lossy. The defining downward flow.S3 (Maintenance Path), S4 (Optimization)
EvictionOut of Tier-NRemoving an entry from a tier to free capacity. May trigger consolidation (writeback) or discard.S3 (Maintenance Path)
PromotionTier-(N+1) to Tier-NMoving data from a slower tier to a faster tier. In GC: objects surviving collection are promoted to an older generation.S4 (Optimization), S5 (Colony)
RetrievalQuery across tiersFetching data from a lower tier on a cache miss at a higher tier. Page-fault cascade.S3 (Read Path)
InvalidationCascade from sourceMarking cached entries stale when the source of truth changes. Propagates upward through tiers.S3 (Maintenance Path)
IngestionExternal to tierNew data entering the system at a physics-appropriate tier.S3 (Write Path)

Ingestion paths:

PathDefinitionEntry Point
Hot ingestionRuntime sources entering the system on the latency-critical path.Tier-0 or Tier-1. Conversation turns, tool outputs, system events, user corrections.
Cold ingestionOffline sources entering the system on the batch path.Tier-2 or Tier-3. Documents, PDFs, conversation history replay, identity seeds.

Collision notes:

Old TermCanonical TermReason
”hot manifold” / “cold manifold”hot ingestion / cold ingestion”Manifold” is a differential geometry term with no CS precedent in this context. “Hot path” and “cold path” are standard in systems engineering.
”culling” (for eviction)evictionStandard cache/HSM term. “Culling” reserved for lifecycle sweep context.
”promotion” (as in promotion_divide)See “delegation spawn” in S2.3”Promotion” is freed for its standard CS meaning: upward tier movement.

The agent lifecycle maps to a hybrid of OS process model and BEAM/OTP actor model. Agents are BEAM processes with structured state.

Canonical NameDefinitionCS SourceUsed In
SpawnCreating a new agent process. Kernel-mediated: the parent expresses intent, the kernel decides. Three subtypes below.BEAM spawn/3, POSIX fork+execS3 (Action Cascade), S5 (Colony Formation)
Renewal spawnSpawn triggered by context exhaustion. Child is a successor — same role, compressed state handoff, fresh context. The parent’s context is full; the child continues the work.Generational GC promotion + process restartS3 (Lifecycle Gates), S5
Specialization spawnSpawn triggered by scope drift. Child is a specialist — role derived from accumulated off-topic work. Parent resets to generalist.Process specialization, actor splittingS5 (Colony Formation)
Delegation spawnSpawn triggered by recurring subtask patterns. Child is a handler — dedicated to the most-pushed work category. Parent zeros that category in the push ledger.Master-worker delegation, Contract Net ProtocolS5 (Colony Formation)
TerminateAgent process ends. Resources released. Evaluation record preserved as tombstone.POSIX exit(), BEAM process terminationS3, S5 (Cull Decision)
State digestCompressed summary of parent’s conversation, injected into child’s system prompt at spawn. Lossy. Max 512 tokens. LLM-generated.Checkpoint + compressionS3 (Lifecycle Gates), S5
Lifecycle sweepPeriodic maintenance: decay scores, terminate low-fitness agents, blend evaluation data.GC collection cycleS3 (Maintenance Path)

Collision notes:

Old TermCanonical TermReason
senescence_divide / “senescence division”renewal spawn”Senescence” is biology (cellular aging). The CS equivalent is context exhaustion triggering process restart. “Renewal” connects to CDR Theorem 3 (Finite Optimal Lifespan).
pollution_divide / “pollution division”specialization spawn”Pollution” is a novel coinage with negative connotation. The actual mechanism is scope drift triggering specialization.
promotion_dividedelegation spawn”Promotion” in CS means moving data UP a hierarchy. The old promotion_divide moves a subtask pattern DOWN to a dedicated handler — the opposite direction. “Delegation” is the standard term.
”cell division” / “division” / “birth”spawnStandard BEAM term.
”death”terminationOS/BEAM standard.
”briefing”state digestCS-adjacent: checkpoint + compression.

Canonical NameDefinitionMutabilityUsed In
ReceptorPublic identity. What the agent handles. The only thing the router sees. Contains embedding vector + capability descriptor. Created at spawn, never modified.ImmutableS3, S5 (Routing)
CharterPrivate operating instructions composed into the system prompt. How the agent handles its work.Immutable for agent lifetimeS3, S8
ScopeLearned semantic fingerprint of work handled. Embedding centroid + exemplars. Drives affinity calculations and specialization pressure. Updated per-message.MutableS3, S5
AffinityHow well a new input matches an agent’s scope or receptor. Scalar [0,1]. Cosine similarity.Computed per-inputS3 (Action Cascade)
Scope driftHow much absorbing a new input would distort an agent’s scope. Scalar [0,1]. Computed as (1 - affinity) * establishment. Drives the spawn decision.Computed per-inputS3 (Action Cascade), S5
EstablishmentDegree to which an agent has built a stable scope fingerprint. Ramps 0 to 1 over the first ~5 ticks. New agents do not spawn immediately.Mutable (monotonic ramp)S3 (Lifecycle Gates)
TickMonotonic counter of completed work cycles (chat rounds).Mutable (monotonic increment)S3
FitnessComposite score [0,1] from four evaluation tiers. Used for routing weight and spawn eligibility.MutableS3, S5 (Cull Decision)

Collision notes:

Old TermCanonical TermReason
”fit”affinity”Fit” is ambiguous (fitness? goodness of fit?). “Affinity” is standard in scheduling (processor affinity) and already used in the architecture docs.
”pollution” (as a pressure metric)scope drift”Pollution” is a novel metaphor. “Scope drift” is descriptive and maps to “distributional drift” from information theory.
state.score (legacy field)fitnessSingle fitness value from the evaluation system. No parallel scoring.

Why “receptor” and “charter” are retained from biology: No CS term captures these concepts without distortion. “Service descriptor” (SOA) and “capability advertisement” (MAS) exist for receptor but are awkward and miss the binding-affinity semantics. “Agent policy” (RL) and “behavior specification” (OTP) exist for charter but are overloaded. Both terms are precise, collision-free, and well-established in the codebase.


The four-action decision cascade for message handling, in priority order (cheapest first).

Canonical NameDefinitionCostCS SourceUsed In
RouteDelegate to a better-fit sibling or child agent. Checked first.CheapestMessage routing (actor model), load balancingS3, S5
PushCreate an ephemeral sub-task frame. Cheaper than spawning.MediumStack push, call frameS3
HandleProcess the message directly in current context. Default action.StandardOTP handle_call, handle_cast, handle_infoS3
SpawnCreate a child agent. Three subtypes: renewal, specialization, delegation.Most expensiveBEAM spawn, POSIX forkS3, S5

Cost model terms:

TermFormulaOld Term
Handle costexec_cost + drift_weight * scope_driftself_cost
Spawn costspawn_overhead - route_creditfork_cost
Delegation pressureMax count in the push ledger. When this exceeds threshold, delegation spawn triggers.promotion_pressure

Collision note: The action :self is renamed to :handle. :self as an action name is confusing — “self” usually means “the agent itself” (reflexive pronoun), not “handle this message.” :handle follows the OTP convention.


Canonical NameDefinitionUsed In
FitnessComposite score [0,1] from four evaluation tiers. Used for routing weight and spawn eligibility.S3, S5
Evaluation tiersFour-tier assessment: (1) Mechanical health, (2) Requester satisfaction, (3) Consensus sampling, (4) Activity decay.S3
DecayExponential reduction in fitness over time without activity. Single unified mechanism — applied during lifecycle sweeps.S3 (Maintenance Path)
TombstonePreserved evaluation record of a terminated agent. Enables clonal resurrection.S3, S5

Collision note (C7 — dual decay): The legacy codebase had two parallel decay mechanisms: a multiplicative score decay in lifecycle.ex (0.95x per sweep) and a time-based exponential activity decay in evaluation.ex. These are unified into a single decay mechanism applied during lifecycle sweeps. One decay source, one application point.


2.7 CDR Theorems — Information-Theoretic Foundation

Section titled “2.7 CDR Theorems — Information-Theoretic Foundation”

CDR is the brand name for the three-theorem framework. Individual theorems have proper names below. CDR remains the shorthand when referencing the framework as a whole. Usage: “the CDR framework,” “CDR properties,” “CDR Theorem 1 (Irreversible Compression).”

Canonical NameFormal StatementOperational MeaningUsed In
Irreversible Compression (Theorem 1)Bounded-state processor receiving stream with h > 0 entropy rate reaches threshold T* beyond which unanswerable queries accumulate monotonically and are irrecoverable.Memory loss is inevitable. The set of questions the system cannot answer grows and never shrinks. No reorganization recovers discarded information. (Data Processing Inequality.)S1 (Theoretical Foundation)
Temporal Partitioning (Theorem 2)For a multi-scale source, a rate-matched partition strictly outperforms any monolithic encoder in worst-case distortion. Fragility ratio >= K-fold.Separating fast-changing and slow-changing state into different tiers reduces worst-case information loss. This is WHY the tier hierarchy exists.S1, S3
Finite Optimal Lifespan (Theorem 3)If degradation grows unboundedly (drift or perturbation), the long-run cost has a finite minimizer tau*. Replace when instantaneous degradation equals long-run average cost: D(tau*) = J(tau*).Every compression scheme eventually degrades enough that rebuilding from scratch is cheaper than patching. This is WHY renewal spawn exists.S1, S3, S7
Recursive Necessity (Theorem 4)CDR applies at every hierarchy level. Compressed output of partition i becomes input to partition i+1.The three properties above are not design choices — they recur at every scale (memory, protocol, population).S1

Collision note: The original CDR shorthand names “Compress / Divide / Renew” are retired as theorem names. “Compress” becomes Irreversible Compression (emphasizes irreversibility, the novel claim). “Divide” becomes Temporal Partitioning (standard CS term; “divide” collides with agent division). “Renew” becomes Finite Optimal Lifespan (names the theorem’s actual result). CDR as the framework acronym is unchanged.

Supporting information-theoretic terms (standard, used as-is):

TermMeaning in Synixolis Context
Rate-distortion function R(D)Fundamental tradeoff between compression rate and information loss
Information bottleneckCompression that preserves task-relevant information
Distributional driftChange in the statistical properties of incoming data over time
Perturbation accumulationBuildup of noise/errors in compressed representations
Unanswerable setThe set of queries the system can no longer answer due to compression
Sufficient statisticA compressed representation that preserves all task-relevant information
Transparency conditionRequirement that compression preserves queryability
Hierarchy depth boundK <= log(T/tau_min)/log(r) — maximum useful number of tiers

Canonical NameDefinitionUsed In
KernelThe BEAM/OTP runtime substrate. Manages process scheduling, message passing, fault isolation. The foundation layer.S3, S8
Memory managerThe Synixolis system itself — sits above the kernel, below agents. Manages tier lifecycle, ingestion routing, consolidation scheduling, evaluation.S3
ArchiveShared knowledge store. Agents read; curators write. Hybrid search (semantic + keyword).S3 (Read Path)
CuratorSpecialized agent with elevated archive write permissions. Analyzes agent lifespans, indexes patterns.S5, S8
Telemetry logAppend-only event log (JSONL). Every spawn, termination, routing decision, evaluation. Replayable.S3 (Telemetry), S7b (Fault Isolation)
ColonyThe full population of living Synixolis-conforming nodes operating over a shared or federated memory substrate.S5, S5b (Collective State)
Clonal lineA generational sequence of agents sharing a receptor. Produced by renewal spawns.S5

Collision note (C1 — “Kernel”): The Synix.Kernel Elixir module is a config profile (LLM client, model, persistence settings), not a kernel. Renamed to Synix.Config. “Kernel” is reserved for the infrastructure substrate.

Collision note (C11 — “Archive”): “Archive” as noun (the knowledge store) and verb (to store in the archive) are both standard English. No rename needed; context disambiguates.

Canonical NameDefinitionUsed In
ConsensusAgreement across colony nodes. Requires N/2+1 nodes validating a proposal against their own traces.S5 (Colony), S7 (Self-Modification)
QuorumThe minimum number of nodes required for a colony-level decision. Default: N/2+1 (simple majority).S5, S7b (Fault Isolation)
Promotion (colony context)Moving a validated pattern from node-local conventions to shared colony-wide conventions. Requires cross-node evidence and consensus.S5 (Organizational Memory), S7
Collective stateShared memory and conventions visible to all colony nodes. Replicated log with materialized views. Causal consistency.S5b (Collective State)
HeartbeatLiveness signal. Written to node manifest on each interaction (not on a timer — Synixolis nodes are session-based, not daemon-based).S5c (Node Lifecycle)
Steering parametersHuman-provided constraints within which the colony self-organizes. Includes max colony size, budget, domain constraints, approval gates.S5 (Colony), S8
StateDescriptionCapabilitiesTransition TriggerUsed In
INITIALIZINGNode bootstrapping from parent digest + colony checkpoint.Read-only on all state. No colony participation.Bootstrap complete -> ACTIVES5c
ACTIVENormal operation.Full read/write on local state. Read + promotion pipeline on shared state. Colony participation.Session ends without termination -> DORMANTS5c
DORMANTSession ended, state persisted on disk, no active process.No operations. State is cold but intact.Session resumes -> ACTIVE. Dormancy exceeds threshold -> liveness check.S5c
FENCEDFault isolation in effect.Read-only on shared state. Full local operation. No consensus participation.Investigation clears -> ACTIVE. Confirmed fault -> TERMINATED (with RESET).S5c, S7b
TERMINATEDNode permanently decommissioned.None. State archived. Colony manifest updated.Cull decision via consensus or human directive.S5, S5c

Canonical NameDefinitionDS AnalogueUsed In
Invariant enforcementPer-operation validation: provenance chain checks, schema enforcement, rate limiting, conflict detection. Stateless, deterministic, never learned.Input validation, referential integrity, token bucket rate limitingS7b
Fault signatureA learned detection pattern (metric + threshold + window) compiled from a resolved incident. Carries provenance, expiry, and its own tau*. Undergoes CDR renewal.Anomaly-based intrusion detection ruleS7b
FencingRevoking a node’s write access to shared state while allowing continued local operation. Time-bounded (default 24h). Reversible.Circuit breaker, network partitionS7b
LevelTriggerResponseReversibilityDS Pattern
MONITORAnomaly score elevated, below thresholdIncrease telemetry sampling rate for suspect node/ruleFully reversibleHealth check
FENCEThreshold exceeded or signature matchedRevoke shared-state write access. Node continues local operation.Reversible on investigation clearanceCircuit breaker
ROLLBACKConfirmed contamination of shared stateRevert specific shared entries/conventions to pre-contamination checkpoint.Partially reversible (rolled-back entries archived with REVERTED flag)Checkpoint restore
EXCISEIdentified harmful entries with known blast radiusRemove specific entries. Provenance audit on downstream dependencies.Reversible (excised entries archived with EXCISED flag)Targeted state cleanup
RESETNode confirmed as comprehensively compromisedWipe node state, re-initialize from last known good checkpoint + state digest. Requires human approval.Irreversible for node stateNode restart / failover
ActionRequired Evidence
Raise MONITORAny single node (unilateral)
Raise FENCE2+ nodes flagging the same target, OR 1 node + invariant violation
Raise ROLLBACKN/2+1 nodes agreeing on contamination scope (majority quorum)
Raise RESETHuman approval (out-of-band escalation)

Canonical NameDefinitionUsed In
PipelineDeclarative DAG of stateless transforms. Transcripts to artifacts. The batch processing system.S3, S8
Pipeline specCompiled representation of a pipeline: receptor + stages + evaluation + projection. Bridge between DSL and kernel execution.S3, S8
Transform patternsFour execution shapes: Map (1:1), Group (N:M), Reduce (N:1), Fold (N:1 sequential).S3
Seeded strategyA human-authored pipeline deployed to the kernel. Subject to evolutionary pressure like any other agent.S5

Collision note (C5 — “Pipeline”): “Pipeline” is overloaded across CS. In Synixolis, “pipeline” or “Synix pipeline” when specificity is needed; “pipeline” alone when context disambiguates. Competitor usage is quoted with attribution: “Cognee pipeline,” “LangChain chain,” etc.


2.11 Metabolic Regions (Node-Level Memory Partitions)

Section titled “2.11 Metabolic Regions (Node-Level Memory Partitions)”

The node contract (S3) specifies five metabolic regions, which are the physical partitions within a single node where the four tiers materialize as storage.

RegionTier MappingDecay RateLifecycleUsed In
ephemeral/Tier-0 + transient Tier-1Hard-expire at TTLHours. The only region where true deletion occurs.S3, S8
operational/Tier-1Active decay (default 14 days)Days to weeks. Sprint goals, blockers, observations.S3, S8
structural/Tier-2Review cycle (monthly default)Months to quarters. Architecture, entities, durable patterns.S3, S5b, S8
identity/Tier-3Permanent until explicitly changedMission, values, non-negotiables.S3, S8
glacier/Archived Tier-2/3None (archived)Indexed. Retrieved on demand. Never deleted.S3, S8

Complete index of resolved terminology collisions. Each collision is documented with the conflicting usages and the resolution rationale.

IDTermConflictResolutionDetails
C1KernelArchitecture docs (“BEAM is the kernel”) vs. Synix.Kernel Elixir config module”Kernel” reserved for infrastructure substrate. Config module renamed to Synix.Config.S2.8
C2PromotionGC/HSM: upward tier movement vs. promotion_divide: spawning a handler”Promotion” reserved for upward tier movement. Handler spawn renamed to “delegation spawn.”S2.2, S2.3
C3Tier numberingCDR Level 1/2/3 vs. Ring-0/1/2/3 vs. AAP Layer 0/1/2Single canonical numbering: Tier-0 through Tier-3. All documents converge.S2.1
C4GenerationAAP: increments on identity mutation only vs. division.ex: also increments on spawnAAP definition is canonical. Generation increments on identity mutation ONLY. Code bug — generation should NOT increment on spawn.—
C5PipelineSynix batch system vs. generic CS term vs. competitor usageContext disambiguates. “Synix pipeline” when specificity needed. Quote competitors: “Cognee pipeline.”S2.10
C6Hot pathGC: access frequency vs. ring architecture: “hot manifold” (runtime ingestion)“Hot path” reserved for access frequency (standard systems term). “Hot manifold” renamed to “hot ingestion.”S2.2
C7Decaylifecycle.ex: multiplicative score decay per sweep vs. evaluation.ex: time-based exponential activity decayUnified into single decay mechanism applied during lifecycle sweeps. One source, one application point.S2.6
C8Score vs. Fitnessstate.score (legacy field) vs. evaluation.fitness (composite)state.score deprecated. Single canonical term: fitness.S2.4
C9RenewCDR: rebuild when degradation > rebuild cost vs. kernel: senescence_divide on context overflowCDR Theorem 3 = “Finite Optimal Lifespan” (the theory). “Renewal spawn” (the implementation). Gap noted: kernel triggers on capacity only; CDR also requires drift-triggered renewal.S2.3, S2.7
C10Spawn vs. Divisiondivision.ex module vs. spawn.ex module vs. “cell division” in docs”Spawn” is canonical (BEAM native term). All docs use “spawn.”S2.3
C11ArchiveNoun (knowledge store) vs. verb (to store)No rename needed. Both usages are standard English and unambiguous in context.S2.8
C12Receptor vs. ScopeConfusion about whether these are synonymsComplementary, not synonymous. Receptor = immutable public routing identity (static). Scope = learned semantic fingerprint (dynamic). Both are required. scope_terms (legacy bag-of-words) deprecated, replaced by scope centroid.S2.4

These terms are kept because no CS term captures the same concept without losing precision or adding confusion.

TermOriginWhy Retained
ReceptorBiochemistry (receptor binding)“Service descriptor” (SOA) and “capability advertisement” (MAS) exist but do not capture the binding-affinity semantics. The receptor metaphor (what an agent binds to) is precise and well-established in the codebase.
CharterGovernance (organizational charter)“Agent policy” (RL) and “behavior specification” (OTP) exist but are overloaded. Charter = private operating instructions defining how an agent works. Clear and collision-free.
ColonyBiology (ant colony, bacterial colony)“Population” (evolutionary computation) is the closest CS term but lacks the connotation of emergent collective behavior. “Colony” captures organism-level dynamics.
Clonal lineBiology (clonal reproduction)“Process lineage” does not capture the identity-preservation semantics (same receptor across generations). “Clonal” describes precisely what happens: the receptor is cloned, the context is refreshed.
TombstoneAlready standard CSDatabase tombstone markers, GC tombstones. Not actually a biology-only term.
FitnessEvolutionary biology, but standard CSHolland (1975). fitness function = evaluation function = objective function. Well-established in evolutionary computation.

Old term to canonical term. For grep-and-replace across the codebase and all documents.

Old TermCanonical TermReason
Ring-0 / Ring-1 / Ring-2 / Ring-3Tier-0 / Tier-1 / Tier-2 / Tier-3Avoid OS privilege ring confusion
Level 1 / Level 2 / Level 3 (CDR)Tier-1 / Tier-2 / Tier-3Unify numbering
senescence_dividerenewal spawnConnect to CDR Theorem 3; remove bio term
pollution_dividespecialization spawnRemove negative connotation; descriptive
promotion_dividedelegation spawnFree “promotion” for standard CS meaning
cell division / division / birthspawnStandard BEAM term
hot manifold / cold manifoldhot ingestion / cold ingestionRemove non-CS “manifold”
pollution (pressure metric)scope driftDescriptive, maps to distributional drift
fit (affinity metric)affinityUnambiguous
self_cost / fork_costhandle cost / spawn costMatch action names
:self (action) / :fork (action):handle / :spawnOTP/BEAM convention
promotion_pressuredelegation pressureFree “promotion”
briefingstate digestCS-adjacent (checkpoint + compression)
VMMmemory managerMore general, less OS-specific
archivistcuratorStandard knowledge management term
score (state.score)fitnessUnify with evaluation fitness
death (agent)terminationOS/BEAM standard
system_basecharterAlready deprecated; complete migration
TermDomainWhy Keep
receptorAgent identityNo better CS term (see S2.13)
charterAgent identityNo better CS term (see S2.13)
scopeAgent stateStandard (working set, topic model)
archiveKnowledge storeStandard CS
consolidationDownward data flowStandard (neuroscience + systems)
evictionTier removalStandard (cache, HSM)
retrievalUpward data flowStandard (cache, memory)
invalidationStale markingStandard (cache coherence)
ingestionExternal data entryStandard (ETL, data engineering)
fitnessEvaluation scoreStandard (evolutionary computation)
decayScore reduction over timeStandard (GC, cache, neuroscience)
tombstoneDead agent’s recordStandard (databases, GC)
colonyAgent populationAcceptable non-CS term (see S2.13)
clonal lineGenerational agent sequencePrecise (see S2.13)
telemetryEvent loggingStandard
tickWork cycle counterStandard
establishmentWarm-up rampDescriptive
push ledgerSubtask frequency mapDescriptive
pipelineBatch DAG processingStandard
CDRThree-theorem frameworkEstablished project acronym

How Synixolis vocabulary maps to established terminology across disciplines. For use in papers, outreach, and positioning.

Synixolis TermCS / SystemsCognitive ScienceAgent Memory Papers (2024-2026)
Tier-0 (Execution)L1 cache, register fileWorking memory (Baddeley)In-context memory, main context (MemGPT)
Tier-1 (Session)L2 cache, working setShort-term memory, episodic bufferInside-trial memory, session memory
Tier-2 (Experience)L3 cache, tiered storageEpisodic memory (Tulving)Cross-trial memory, long-term memory
Tier-3 (Identity)Permanent generation (GC), cold storageSemantic memory (Tulving)Parametric memory, core memory
ConsolidationWriteback, demotion, GC promotionSystems consolidation (hippocampus to neocortex)Memory compression, reflection (Park et al.)
EvictionCache eviction, LRU replacementForgetting, decayMemory forgetting, pruning
Renewal spawnProcess restart, checkpoint-restoreMemory reconsolidation(No equivalent — novel)
Scope driftDistributional drift, concept driftSchema drift(Not formalized in agent memory lit)
AffinityCache affinity, processor affinity—Relevance score, similarity
State digestCheckpoint compression, snapshotMemory consolidationSummarization, compression
Irreversible Compression (T1)Data Processing InequalityCatastrophic forgetting, interferencePartially: stability-plasticity dilemma
Temporal Partitioning (T2)Tiered caching, HSMComplementary Learning Systems (CLS)Hierarchical memory (HMT, H-MEM)
Finite Optimal Lifespan (T3)GC generational hypothesis, cache TTLSleep-dependent consolidation cycles(Novel — no equivalent)
FitnessFitness function (GA), objective function—Evaluation score
ColonyPopulation (evolutionary computation)—Multi-agent system
CuratorKnowledge curator, ontology maintainer—Memory manager (Memory-R1)