Memory Architecture

The Nooplex Memory Stack — how ProbOS organizes knowledge from fleeting thoughts to permanent institutional memory.

---

Overview

ProbOS implements a six-layer memory architecture inspired by the Nooplex paper's cognitive ecosystem model. Each layer serves a distinct function in the knowledge lifecycle, with well-defined promotion and demotion paths between layers.

Two cross-cutting systems connect the layers: - Ward Room Bus — lateral knowledge flow between agents at any layer (social fabric) - Dream Consolidation — vertical knowledge promotion from experience to long-term storage (the elevator)

The Stack

┌─────────────────────────────────────────────────┐
│  1. Global Workspace                            │
│     Multi-agent reasoning, DAG orchestration    │
│                                                 │
│  ┌──────────── Ward Room Bus ────────────────┐  │
│  │  Social knowledge fabric (lateral flow)   │  │
│  └───────────────────────────────────────────┘  │
│                                                 │
│  2. Ephemeral Working Memory                    │
│     Agent scratchpads, context windows          │
│                                                 │
│          ↕ Dream Consolidation (elevator)       │
│                                                 │
│  3. Vector Store (Associative Cortex)           │
│     EpisodicMemory, SemanticKnowledgeLayer      │
│                                                 │
│          ↕ Pattern extraction / promotion       │
│                                                 │
│  4a. Formal Models (Constitutional)             │
│     Ontology, Standing Orders, Skill Defs       │
│  4b. Operational Knowledge (Learned)            │
│     Trust records, Hebbian weights, profiles    │
│                                                 │
│          ↕ Publication / archival               │
│                                                 │
│  5a. Institutional Memory (Ship's Records)      │
│     Notebooks, logs, reports, Captain's Log     │
│  5b. Operational State (KnowledgeStore)         │
│     Trust snapshots, routing weights, agent src │
│                                                 │
│          ↕ Federation push/pull                 │
│                                                 │
│  6. Distributed Storage Substrate               │
│     Instance repos, fleet remotes, Nooplex      │
└─────────────────────────────────────────────────┘

Layer Details

Layer 1: Global Workspace

The shared cognitive space where multi-agent reasoning happens. When a Captain's intent enters the system, it becomes a TaskDAG — a directed acyclic graph of cognitive work distributed across agents. The workspace holds the current state of all active reasoning: which agents are thinking about what, what results have been produced, what depends on what.

ProbOS implementation: Runtime DAG orchestration, IntentBus broadcast, consensus voting.

Analogy: The bridge viewscreen during a crisis — everyone sees the same situation, contributes their expertise, and the picture updates in real-time.

Ward Room Bus (Lateral)

Not a layer but a cross-cutting communication fabric. The Ward Room enables lateral knowledge flow between agents at any point in the stack. An agent's insight during Layer 1 reasoning can be shared as a discussion post. A pattern noticed during Layer 3 recall can spark a department-wide conversation. The bus is bidirectional — agents both publish to and learn from Ward Room interactions.

ProbOS implementation: WardRoom (channels, threads, posts, endorsements), @callsign addressing, department broadcasts.

Key property: "Brains are brains" — human and AI participants share the same bus. The Captain is @captain on the Ward Room, not a special external interface.

Layer 2: Ephemeral Working Memory

Each agent's short-term scratchpad. Context windows, intermediate reasoning, tool call results, conversation history for the current cognitive cycle. Discarded after the cycle completes — nothing persists here by default.

ProbOS implementation: LLM context windows, compose_prompt() assembled context, working memory dict in process_natural_language().

Key property: Cheap to create, expensive to lose mid-cycle, zero cost to discard after. The Selective Encoding Gate (AD-433) controls what gets promoted from here to Layer 3.

Dream Consolidation (Elevator)

The vertical promotion mechanism. During dream cycles (offline consolidation), the system: 1. Replays recent episodic memories (Layer 3) 2. Extracts strategy patterns via LLM analysis 3. Updates Hebbian weights and trust records (Layer 4b) 4. Prunes low-value connections 5. Pre-warms routing for the next wake cycle

Dream consolidation is biologically inspired — it mirrors how mammalian brains consolidate short-term memories into long-term storage during sleep. The "elevator" metaphor: experiences ride up from Layer 3 to become permanent knowledge in Layers 4-5.

ProbOS implementation: DreamingEngine (cognitive/dreaming.py), strategy extraction, Hebbian weight consolidation, trust snapshot persistence.

Layer 3: Vector Store (Associative Cortex)

Persistent associative memory. Experiences, semantic embeddings, and contextual knowledge stored as vectors for similarity-based retrieval. This is where "what happened" lives — not interpreted or distilled, just recorded.

ProbOS implementation: - EpisodicMemory — autobiographical episodes (what happened, who was involved, outcome, emotional valence). Sovereign per-agent shards. - SemanticKnowledgeLayer — five ChromaDB collections (agent designs, skills, episodes, context, artifacts). - CognitiveJournal — structured per-agent decision records with traceability chains (AD-431/432).

Key property: Content-addressable retrieval. "Find me experiences similar to this situation" — not keyword search, but meaning-based recall.

Layer 4: Structured Knowledge

Split into two sub-layers reflecting fundamentally different knowledge types:

4a. Formal Models (Constitutional) — Authored, versioned, human-approved. Changes require deliberate design decisions. - Vessel Ontology (AD-429) — formal concept hierarchy across 8 domains - Standing Orders (AD-339) — 4-tier behavioral constitution - Skill definitions — procedural templates - Alert condition definitions

4b. Operational Knowledge (Learned) — Emergent from experience. Changes continuously through interaction. - Trust records — per-agent Bayesian trust scores - Hebbian weights — connection strengths between agents - Agent personality profiles — Big Five traits evolved through experience - Routing preferences — learned intent-to-agent mappings

Key property: 4a is "the law" — stable, deliberate, version-controlled. 4b is "wisdom" — emergent, continuously updated, reflects the system's lived experience. Both inform behavior, but 4a constrains while 4b guides.

Layer 5: Persistent Storage

Split into two sub-layers reflecting different persistence models:

5a. Institutional Memory (Ship's Records) — AD-434. The written story of the crew. Git-backed documents with YAML frontmatter providing systematic provenance (author, classification, status, topic, tags, timestamps). - Captain's Log — append-only command narrative - Agent Notebooks — per-callsign research notes, observations, working documents - Reports — formal published findings (precedent documents, case law) - Operations — duty logs, watch reports, incident records - Manuals — procedures, checklists, reference materials

5b. Operational State (KnowledgeStore) — System checkpoint persistence. NOT a shared knowledge library despite the name. - Trust snapshots — serialized trust network state - Routing weights — Hebbian connection weights - Agent source code — self-designed agent configurations - QA reports — system quality assessment results - Extracted strategies — dream cycle outputs

Key distinction: 5a is authored knowledge — agents and humans write documents that other agents can read and learn from. 5b is system state — machine-generated checkpoints for warm boot and recovery. An agent reads 5a like a library; an agent restores 5b like loading a save file.

Layer 6: Distributed Storage Substrate

The physical persistence and federation layer. Everything above eventually lands here — Git repositories, file systems, federation remotes.

ProbOS implementation: - Instance-level: local Git repos (KnowledgeStore, Ship's Records) - Fleet-level: Git remotes shared between federated ProbOS instances - Nooplex-level: cross-fleet knowledge exchange (future)

Key property: Federation happens here. Git's distributed model means every instance has a complete copy, merge conflicts are resolvable, and history is immutable. Knowledge flows between ships through push/pull, not real-time replication.

Three Knowledge Tiers

A simplified view for practical reasoning about where knowledge belongs:

Tier	Name	Store	Nature	Example
1	Experience	EpisodicMemory	Raw autobiographical episodes	"I scanned the codebase and found 3 security issues"
2	Records	Ship's Records (AD-434)	Structured documents with provenance	Security audit report, engineering log, research notebook
3	Operational State	KnowledgeStore	System checkpoints	Trust snapshots, routing weights, agent source code

Promotion paths: - Tier 1 → Tier 2: Agent writes up findings from episodes into a notebook or report - Tier 1 → Tier 3: Dream consolidation extracts patterns, persists weights - Tier 2 → Tier 3: KnowledgeStore bridge indexes Ship's Records documents

SECI Knowledge Creation Model

ProbOS's memory architecture maps to Nonaka & Takeuchi's (1995) SECI knowledge creation cycle:

Phase	Description	ProbOS Implementation	Status
Socialization	Tacit → Tacit. Knowledge shared through shared experience	Ward Room discussions, agent-to-agent conversation	Implemented
Externalization	Tacit → Explicit. Knowledge articulated into formal concepts	Ship's Records — agents write findings into notebooks/reports	AD-434 (planned)
Combination	Explicit → Explicit. Formal knowledge recombined into new insights	Ship's Records cross-referencing, Oracle queries across tiers	AD-434 + Oracle (planned)
Internalization	Explicit → Tacit. Formal knowledge absorbed through practice	SemanticKnowledgeLayer recall → agent behavior change	Partial (data is operational state, not authored knowledge)

Gap analysis: Before AD-434, ProbOS had Socialization (Ward Room) and partial Internalization (semantic recall), but Externalization and Combination were missing entirely. The duty-output pipeline (agents produce professional records as part of duties) had no place to store results — findings evaporated after each cognitive cycle.

Duty-Output Pipeline

Every crew member produces professional records as part of their duties. Ship's Records provides the destination:

Role	Duty Output	Records Location
Medical Officer	Patient assessments, crew fitness	notebooks/bones/, operations/medical/
Counselor	Crew evaluations, cognitive health	notebooks/troi/, reports/
Engineer	System scans, modification logs	notebooks/laforge/, operations/engineering/
Security	Threat assessments, incident reports	notebooks/worf/, operations/security/
Science	Research findings, analysis	notebooks/number-one/, reports/
Operations	System status, operator rounds	notebooks/obrien/, operations/
Scout	External research, tool evaluations	notebooks/wesley/, reports/

Nooplex Paper Alignment

The "Shared Memory" principle from "The Nooplex: A Planetary Cognitive Ecosystem" (Galliher, 2026) was initially listed as well-covered by ProbOS, but deeper analysis (AD-434 research) revealed the shared knowledge fabric is not implemented:

• KnowledgeStore was intended to be the "shared library" but evolved into operational state persistence
• No agent writes knowledge to it — only system processes (dream consolidation, QA, trust snapshots)
• Dream consolidation modifies in-memory weights but does not promote distilled insights to a shared corpus

AD-434 (Ship's Records) is designed to be the actual implementation of the Nooplex shared knowledge fabric: - Captain's Log (human contributes) → agent notebooks (agents amplify & document) → Ward Room discussion (human refines) → dream consolidation → notebook entries (substrate evolves) - This four-phase cycle maps to the SECI knowledge creation model - It also implements the Nooplex paper's "Human-Agent Knowledge Feedback Loop" (Gap #10)

Related Documents

Document	Scope
Knowledge Layer	KnowledgeStore and SemanticKnowledgeLayer implementation details
Cognitive Architecture	Dream consolidation, strategy extraction
Experience Layer	EpisodicMemory, Selective Encoding Gate
AD-434 (Ship's Records)	Detailed design for institutional memory
Roadmap — Research	Nooplex paper alignment gaps

Source Files

File	Layer	Purpose
knowledge/store.py	5b	Git-backed operational state persistence
knowledge/semantic.py	3	SemanticKnowledgeLayer (5 ChromaDB collections)
cognitive/episodic_memory.py	3	Autobiographical episode storage
cognitive/dreaming.py	Elevator	Dream consolidation engine
cognitive/strategy_extraction.py	Elevator	Pattern extraction from episodes
cognitive/cognitive_journal.py	3	Per-agent decision records
cognitive/encoding_gate.py	2→3	Selective encoding (what gets remembered)
ward_room.py	Bus	Social communication fabric
consensus/trust.py	4b	Bayesian trust records
mesh/routing.py	4b	Hebbian connection weights