Skip to content

Composable Cognition — Cognitive Organs, the Spine, and the Two-Scale Nervous System

Status: Foundational concept (proposed). Date: 2026-06-18. Pilot: the Attention organ — epic #975, attention-architecture.md, attention-research.md. AD ceiling at authoring: AD-1026 landed; AD-1027 reserved (#973); AD-1028..1032 filed (#976–#980). New ADs here open at AD-1033.

Thesis. A ProbOS cognitive agent is an organism: a spine plus a set of cognitive organs. The spine is the agent's central nervous system (synchronous, integrated, private). The mesh is the ship's nervous system (asynchronous, governed, federated). The same pattern — a connective substrate linking cognitive units — appears at both scales with scale-appropriate properties. Society of Mind on the inside; the Nooplex on the outside; the same fractal at every level.

1. Why this concept

ProbOS began with agents that were services for the OS. It now runs a crew — a civilization of independent brains. The open question is no longer "what services does the OS offer?" but "how is a brain composed?" Today a CognitiveAgent is a monolith with helpers and a fixed prompt-assembly pipeline. The recall arc (BF-630→632) and the attention work (#975) showed the cost of that: capabilities that should be distinct faculties (attention, memory, perception, valuation) are tangled into one method with no contract, no composition, and no introspection.

The proposal: make Cognitive Organ a first-class concept and compose brains from organs connected by a spine. This is not a rewrite — ProbOS already has latent organs (see §5); we are naming an emergent structure and giving it a contract.

2. The model

2.1 Cognitive Organ — a child component, not an agent

A Cognitive Organ is a bounded cognitive faculty that a CognitiveAgent is composed of. It is not a mesh agent. The five-part test (a capability is an organ only if it has all of these — otherwise it's a plain module/helper):

  1. A distinct cognitive function — it does one identifiable thing a mind does (attend, remember, perceive, value/feel, monitor-self, narrate, reflect).
  2. Persistent state across cognitive cycles — it carries state between turns (not a pure function).
  3. The perceive → decide → act cognitive-cycle shape.
  4. 1:1 ownership — intrinsic to one agent (personal), not a shared service.
  5. Introspectability — its contribution to behavior is auditable.

Identity & lifecycle (the load-bearing discipline): - Child, not peer. Identity is derived and namespaced under the parent (e.g. {parent_id}.attention). An organ is not in the agent registry, not addressable on the mesh, and has no independent trust score, vote, or consensus standing. It exists only as part of its parent. - Born with the parent, dies with the parent. Constructed inside the parent's birth; torn down inside the parent's teardown. No independent spawn or death. - Shape, not classification. It shares the agent shape (and the Society-of-Mind lineage, where "organs" and "agents" are the same thing), but ProbOS keeps the vocabulary crisp: organs are components; agents are mesh peers. We do not register an AttentionAgent; we compose an AttentionFaculty organ. This avoids the confusion other agent harnesses invite by overloading "agent." - Deterministic by default. Most organs run every cognitive cycle, before the expensive LLM call; their work is arithmetic/structural, not reasoning. (Some organs — reflection, narration — may invoke the LLM; that is the exception, not the rule.)

2.2 The Spine — the agent's central nervous system

The spine is the in-process connective backbone the agent provides. It does three things: - Composition — holds the organ registry ("which organs compose this brain") and attaches/detaches organs at birth/death. - Cycle — drives the perceive → decide → act cognitive cycle across the organs. - Signaling — an in-process channel by which organs influence one another (the valuation organ raises arousal → the attention organ narrows; the attention organ selects memories → the narration organ phrases them).

The spine is synchronous, in-process, ungoverned, and private. You do not hold a quorum vote between your own thalamus and hippocampus, and you do not put a network round-trip between two of your own organs.

2.3 The Mesh — the ship's nervous system

The mesh (the intent bus + gossip + routing) is the ship's nervous system: it connects brains (cognitive agents). It is asynchronous, governed (consensus/trust), and federated (sovereign peers).

2.4 Same pattern, scale-appropriate properties

The fractal is real, but each scale gets its own substrate. Conflating them is the failure mode (a synchronous unified mesh becomes a central scheduler; an async governed spine puts a network hop + consensus inside a single mind).

Spine (agent NS) Mesh (ship NS)
Coupling Synchronous, in-process Asynchronous (pub/sub, gossip)
Governance None (organs aren't sovereign) Consensus / trust / quorum
Topology Integrated — one self Federated — sovereign peers
Boundary private to the agent the ship-wide fabric
Biology central nervous system social/peripheral signaling between organisms

A familiar framing: the mesh is a microservices architecture — autonomous, networked, governed services (ProbOS's intent bus + trust/consensus is, in effect, a cognitive service mesh). The spine is a modular monolith within each service — in-process, synchronous components, not independently deployable, not network-isolated. Same "bounded components, composition over monolith" philosophy; the substrate differs by scale. Calling an organ a microservice would wrongly imply a network hop + independent deployment (the over-engineering trap): organs are in-process modules; agents are the services.

2.5 The single governed boundary (sovereignty preserved)

An agent's nervous system connects to the ship's nervous system through one controlled gateway — the agent's existing mesh interface. Organs do not independently chatter on the mesh; the agent mediates. Exogenous signals (mentions, alerts, camera-change, gossip) enter as an owned inlet that updates organ state between turns. The CNS reaches the world through defined pathways, not every neuron wired to the outside — same discipline, and it is how AD-397 sovereignty is kept.

2.6 The self (Minsky's "no central self," answered)

A society of organs needs something that makes it one mind rather than a swarm. That is the spine plus the agent's instructions (its constitution). Not a homunculus — a connective backbone and an identity. "Who's in charge?" — nobody and the spine, simultaneously.

2.7 Brains within brains (recursion — allow, don't require)

An organ may itself be a composed brain (its own sub-spine + sub-organs) — e.g. a "memory organ" composed of episodic + working + consolidation sub-organs. The architecture permits this recursion but keeps it shallow (1–2 levels) to start. Most organs are leaves; earn depth where a faculty genuinely is a sub-society.

Organs, not neurons (hard guardrail). The granularity stops at the organ. We compose bounded faculties, not neuron-level sub-agents — decomposing an organ into a swarm of tiny "neurons" is over-engineering with no payoff at this scale. The five-part test (§2.1) is the floor: if it isn't a distinct, stateful, introspectable faculty, it's a plain module — not an organ, and never a neuron-agent.

3. The body plan (brain-region map) — and ProbOS's latent organs

The organs aren't arbitrary; they map to a real cognitive architecture, and ProbOS already has seeds of most of them:

Organ Brain analogue ProbOS today (proto-organ)
Deliberation cortex the LLM call (System-2) — the one expensive reasoning core organs surround
Attention / gating thalamus AttentionFaculty (#977, being built)
Memory hippocampus working memory (AD-573) + the episodic shard
Valuation / affect amygdala cognitive zones / arousal (AD-588) + affective salience (#908)
Economic judgment orbitofrontal cortex EconomicJudgmentOrgan (AD-1322, opt-in, default OFF): weighs what an agentic turn spends against what the work is worth, driven per step through an explicit inner-loop hook; informs only, distinct from valuation/affect
Metacognition / self-monitoring prefrontal self-monitoring (AD-504), confab guard (AD-592), source attribution (AD-568)
Interoception / body sense insula avatar self-observation (AD-722)
Consolidation / reflection default-mode network dreaming → becoming a personal organ (decided; AD-1035) so agents dream independently
Perception sensory cortices the sensorium taxonomy (PROPRIO / INTERO / EXTERO) — already organ-flavored

The sensorium layering is the strongest tell that ProbOS has been drifting toward organ composition without naming it.

4. Prior art absorbed

The closest precedents are academic cognitive architectures and a few implementations. None does ProbOS's exact synthesis (deterministic organs as lifecycle-bound child components of an LLM agent, connected by a synchronous spine, inside an async governed mesh of such agents) — so this is a novel synthesis on well-established shoulders. Patterns absorbed; no code vendoring (license-clean).

Cognitive science / architectures (the canon): - Society of Mind (Minsky, 1986) — mind = society of agents in agencies; no central self. The conceptual root. - Modularity of Mind (Fodor, 1983) — encapsulated, domain-specific modules + a central system. → organ boundaries should be crisp; the spine is the central system. - ACT-R (Anderson) — the strongest precedent. A mind composed of modules (declarative, procedural, goal, perceptual-motor), each exposing a buffer, coordinated by a central production system. → organ = module, spine = coordinator, the organ's output = its buffer/bid. - Soar (Laird, Newell, Rosenbloom) — working memory + productions; impasse-driven subgoaling. → deliberate attention shifts on impasse. - Common Model of Cognition / Standard Model of the Mind (Laird, Lebiere, Rosenbloom, 2017) — a consensus modular "body plan." → an agent type composes a canonical organ set; differentiation is compositional. - LIDA (Franklin) — computational Global Workspace: attention codelets compete each cognitive cycle to bring coalitions into the workspace, then broadcast. → the cycle model for the spine + attention organ. - CoALA — Cognitive Architectures for Language Agents (Sumers, Yao, Narasimhan, Griffiths; TMLR 2024; arXiv:2309.02427) — the modern bridge. Frames an LLM agent as modular memory components + a structured action space + a generalized decision procedure. → exactly our organs + spine with the LLM as the deliberation core; validates the approach in the LLM era. - Generative Agents (Park et al., 2023) — memory/reflection/planning modules; retrieval = recency + importance + relevance. → feeds the salience organ.

Implementations (GitHub — patterns absorbed, no vendoring): - ROS 2 (ros2/ros2, Apache-2.0; "a meta operating system for robots") — nodes + topics are a robot's nervous system: many small components coordinated by a pub/sub graph. The canonical "nervous system = pub/sub graph of components." → our mesh is the ROS-graph at the ship scale; we keep the spine synchronous (a robot's reflex arcs aren't all on the async bus). Validates the metaphor. - OpenCog AtomSpace / Hyperon (opencog/atomspace) — "several dozen modules" / MindAgents operating over a shared in-RAM hypergraph; described as "kind-of-like an OS kernel." The cognitive-synergy precedent: specialized processes (organs) over a shared substrate. → organs cooperate via shared state (working memory / the spine's signal channel); the substrate-as-kernel framing. - Letta (formerly MemGPT) (letta-ai/letta, Apache-2.0) — stateful agents with self-editing memory ("memory blocks") and OS-style virtual context management. The memory-organ precedent: a persistent, self-managing memory faculty. → the memory organ owns and edits its own state.

5. Mapping onto ProbOS (you're naming a latent structure)

  • CognitiveAgent._run_cognitive_lifecycle (perceive→decide→act) is the proto-spine today — it already drives the cycle and calls the faculties. Formalizing the spine makes organ attachment (composition) and intra-organ signaling explicit.
  • The intent bus is the mesh (ship nervous system) — already exists.
  • The sensorium registry (PROPRIO/INTERO/EXTERO) is the proto-organ catalog; the AttentionBid catalog (#976) generalizes it.
  • Working memory (AD-573) is the clearest existing personal organ; the episodic store is a shared faculty with per-agent shards — the canonical "personal organ over a shared substrate."

6. Personal organ ⇄ shared ship faculty (a first-class design axis)

A capability can be provided as a personal organ (paired, sovereign, intimate, synchronous) or a shared ship faculty (mesh service, governed, single-source): - Personal organs: attention, working memory, valuation/affect, self-monitoring, interoception, narration. - Shared faculties: the Git knowledge base, event log, consensus, trust network, the episodic store (with per-agent shards), the model router. - Rule: personal when it's intimate to one mind's identity/sovereignty and needs private synchronous access; shared when it's ship-level truth, needs governance, or benefits from a single source. Some capabilities are both (a personal working-memory organ over a shared episodic store). Decided (2026-06-18): dreaming becomes a personal organ (AD-1035) — each agent dreams independently over its own episodic shard, rather than a single shared runtime consolidation pass. This unlocks per-agent dream cadence (some agents dream more than others), independent dream exploration, and the agent interpreting its own dreams (AD-980/#909 becomes an organ capability). The episodic store stays shared (per-agent shards); the dreaming faculty becomes personal.

7. Realization (the attention organ is the pilot — and it carries the foundation)

The epic #975 is scoped so the attention organ establishes the whole pattern. Build order (dependency, not AD-number order):

  1. AD-1028 (#976) — ContextAssembler seam + AttentionBid + global token budget (behavior-preserving, default-OFF). Context mechanics.
  2. AD-1033 (new) — CognitiveOrgan protocol: the five-part contract + base class + the child-identity / born-with-parent / dies-with-parent lifecycle helpers. Pure interface; the foundation.
  3. AD-1034 (new) — CognitiveSpine (thin): organ registry/composition + the cognitive-cycle driver + the in-process intra-organ signal channel + the single governed mesh boundary. Behavior-preserving (byte-identical with zero organs registered); the CognitiveAgent lifecycle is refactored to host it. The backbone.
  4. AD-1029 (#977) — AttentionFaculty: the first organ, conforming to the AD-1033 contract, attached to the AD-1034 spine, driving the AD-1028 assembler. The pilot that proves the pattern.
  5. AD-1030 (#978) — salience scoring · AD-1031 (#979) — camera as a bid (closes #973) · AD-1032 (#980) — exogenous interrupts & arousal.

8. Non-goals & guardrails

  • Not a rewrite — and non-disruptive (load-bearing constraint). Concept + seam + pilot + retroactive classification of the proto-organs (name them without rewriting) + opportunistic migration. Strangler-fig. Every step is behavior-preserving (default-OFF / zero-organ byte-identical); moving the existing cognitive agents to this architecture must never disrupt or break them. The goal is a more elegant architecture that scales — not a big-bang rebuild.
  • Don't metastasize. The five-part test gates what becomes an organ; helpers stay helpers. Resist "everything is now an organ." Organs, not neurons.
  • Don't network the spine. The spine is synchronous, in-process, ungoverned.
  • Behavior-preserving foundations. AD-1033/1034 ship byte-identical (default-OFF / zero-organ); adaptivity comes later and gated.
  • Deterministic-by-default organs — no consensus/trust/LLM for internal calls.

9. Performance & scaling

Guiding principle (2026-06-18): optimize shared resources where possible; preserve autonomy where it provides an advantage. The organs and agents are all one connected system on the ship — share the platform, keep the decisions sovereign. Two cost regimes, handled differently:

Hot-path organs (per-turn) — perf-neutral-to-positive

Attention, working memory, valuation and perception-gating run every cognitive cycle, before the LLM call, but are cheap by construction: deterministic, bounded, and no LLM or network/mesh call on the synchronous cycle. Composing K organs is K cheap calls + bid scoring (arithmetic over a bounded candidate set) — sub-millisecond against a multi-second LLM call. The global token budget (AD-1028) is a net win (smaller prompts ⇒ cheaper, faster calls). Two rules preserve this: reuse the embeddings recall already computes (don't embed every bid fresh), and enforce no await on a bus/network call in the cycle path (the discipline-erosion guard, asserted in AD-1034).

Background organs (idle) — the real scaling concern

Dreaming / reflection / interpretation run during idle and may use the LLM. Naive per-agent execution multiplies the expensive part (a 50-agent crew → up to 50× dreaming LLM cost) with thundering-herd + store-contention risk. Resolution — personal faculty, shared bounded executor (the microservice-with-shared-platform pattern): - Faculty is personal — each agent decides what/when/how-much to dream over its own shard (per-agent cadence; some dream more than others). - Execution is shared & governed — evolve the existing idle/dreaming scheduler into a background-cognition scheduler: a shared ship faculty that dispatches all organs' background work through a bounded worker pool with a global concurrency cap and jitter — not N×K per-organ timers. Same serialization as today's single pass, with per-agent independence on top. - Entrance gate — an organ runs only with enough new salient material since last time (cf. AD-670); idle agents don't pay. - Cheaper tiers + offline — bulk consolidation on fast/local tiers, deep tier only for genuine synthesis; idle-time work, so it never touches per-turn latency. - Promote shared-worthy outputs — genuinely cross-agent results (extracted procedures, anti-patterns) go to the shared knowledge base, so N dreamers don't re-derive the same procedure 50×.

Footprint & measurement

Keep organ state lean (spill to the shared store; no big per-organ buffers); drive background work from the shared scheduler, not per-organ asyncio loops. The behavior-preserving / default-OFF migration lets us benchmark before/after each organ lands — per-turn overhead (target sub-ms) and aggregate background cost (LLM calls/hour); a regression backs out the flag.

10. Decisions (2026-06-18)

  1. Design-Principle elevation — DONE. "Composable cognition (agents are organisms)" is now Design Principle #12 in docs/development/architecture-reference.md, with the cognitive-organ layer noted in the Agent Classification Framework. We elevated now rather than after the pilot — the Captain judged it the right foundation (and an early miss not to have started here), under the hard constraint that migration stays non-disruptive.
  2. Dreaming becomes a personal organ — DECIDED (AD-1035). Agents dream independently over their own shard; see §6. Unlocks dream interpretation, independent exploration, and per-agent cadence.
  3. Recursion depth — capped at the organ level. Organs, not neurons (§2.7); 1–2 levels only, and only where a faculty genuinely is a sub-society.