Context Architecture
Memory is stored; context is activated. A stored memory becomes context only when it is selected into a Cognitive Frame.
Memory answers “what has Lucid retained?” — the durable substrate. Context answers “what matters right now?” — the active operating field a reasoning cycle works inside. It is not a database, and not a prompt blob.
This is the engineering complement to the Epistemic Field Model: EFM describes the structure of epistemic space; Context Architecture describes which slice of that space is activated for a given cycle. Grounded in Global Workspace Theory — the lit stage of working memory, with a spotlight deciding what enters.
Context is distributed across layers, each responsible for a different kind of situatedness. Every layer maps to an existing or scheduled Lucid component — no new stores are introduced. The value is that context selection becomes inspectable rather than buried in a prompt.
Environmental
signals, events, documents, tool outputs, prompts
Signal Scout → Signal Gate
Task
active intent, definition of done, constraints, scope
Cognitive Frame (type=context)
Historical
prior decisions, past cycles, rejected options
Lucid Memory (L1 + cycle_memory)
Domain
concepts, taxonomies, ontologies, domain rules
Temporal Graph Memory (Lucid Atlas)
Human / Social
stakeholders, decision rights, role, sensitivity
Human Co-Agency; Decision actors
Epistemic
assumptions, uncertainty, confidence, contradictions
EFM; assumption_history; stance_tags
Procedural
methods, skills, workflows, operating patterns
ACE / Capability Forge / skills
Artifact
specs, drafts, derivatives, diagrams, code
Editorial / Distribution; drafts
Roadmap
current priorities, milestones, capability gaps
BACKLOG.md · Notion Roadmap
Model / Tool
available models, tools, limits, permissions
Capability Forge registry
The input contract to a single Divergent–Convergent Reasoning cycle.
A Cognitive Frame is the bounded, typed working state for one DCR cycle: what Lucid is reasoning about right now, under which assumptions, constraints, and goals, with which context admitted.
It requires no schema change — a Frame is stored as an epistemic_objects row of type='context' and links to its trigger and admitted items through the existing relationship graph. Following the limited-capacity principle, each layer holds references and activation scores, not full content — the Frame must be small enough to assemble on every cycle.
Its required_depth field is the seam where CAML’s Reflective Depth enters.
The Loom is the function that produces one Cognitive Frame from a trigger and an active goal — and logs why each candidate item was admitted, summarized, or dropped. It is not a new service: Lucid already performs this work implicitly in the Research Curator and the Memory Retrieval Broker. The Loom names that work, types its output, and makes it inspectable.
It reuses existing scoring rather than reinventing it — the Retrieval Broker’s composite and the Signal Gate’s novelty/recurrence/goal-relevance. Each activation decision — include · summarize · ignore · escalate · ask_human — is itself a decision: logged, evaluated after the cycle, and calibrated with Capability Pressure deltas.
This yields a Context Quality Score (CQS), analogous to DQS and TQS — falsifiable dimensions only. Subjective dimensions wait for a feedback path that can prove them wrong.
DCR — the Frame is the input contract to a cycle; the Loom assembles it; the cycle reasons inside it. Context activation situates DCR, it does not replace it.
EFM — the Frame is the activated slice of epistemic space. CAML — its modulation variables are the Loom’s knobs (caution, ambiguity tolerance, reflective depth). ACE — owns the procedural and model/tool context layers.
Only two new concepts are minted here — the Context Loom and the Cognitive Frame — plus the Context Quality Score. Everything else is a renaming of components Lucid already has.
The “Lucid Harness” is an engineering label for the existing runtime, not a seventh theory concept.