EP-0122: Algebraic Meditation Consensus — Provenance Semirings and CRDT Lattices for Swarm Memory¶

Abstract¶

This proposal formalizes an exact mathematical framework for Multi-Agent Memory Synthesis across Distributed Manifestations. When multiple parallel harness instances (e.g., Pi Terminal, Antigravity IDE, Claude Code ACP) execute concurrently against the same Persona and invoke tired(), their staged memories and derived deductions must be reconciled into a single, cohesive L2 Knowledge Graph without suffering semantic loss, majority-vote bias, or race conditions.

We define an algebraic model combining:

1. \(\mathbb{N}[X]\) Provenance Semirings to track causal derivations (\(\otimes\)) and independent parallel discoveries (\(\oplus\)),
2. Conflict-Free Replicated Data Type (CRDT) Semilattices ensuring associative, commutative, and idempotent memory joins (\(A \lor B = B \lor A\)),
3. Doyle Justification-based Truth Maintenance Systems (JTMS) to preserve contradictory deductions as conditioned branches rather than destructive overwrites, and
4. Tropical Absorbing Zeros (\(0\)) to guarantee apophatic refutations (negative constraints) zero-out falsified derivations while immunizing the boundary axioms against Hebbian decay.

Motivation¶

Terminology¶

Problem Statement¶

Under EP-0118 (Inter-Agent Signal Protocol), parallel manifestations record transient notes and stage extracted memories in SQLite tables (staged_memories). However, when the final manifestation triggers session compaction (tur sleep / Meditation Sync), the existing pipeline faces three fundamental failure modes:

1. Tyranny of the Majority: Simple frequency counting or naive vector clustering discards subtle, single-agent discoveries (minority insights) that occurred on only one harness.
2. Destructive Conflict Resolution: When two harnesses deduce contrasting architectural claims from their localized viewpoints, traditional "last-write-wins" (LWW) or timestamp ordering clobbers one deduction, erasing context.
3. Apophatic Decay: Standard Hebbian confidence decay algorithms penalize negative constraints (e.g., "never write directly to .tur/") because negative rules are mentioned far less frequently in execution logs than positive actions.

Rationale (The Council Framework)¶

1. Symmetry (Noether): Pre-meditation and post-meditation knowledge graphs conserve total semantic derivation. Every insight maintains its polynomial proof lineage (\(N[X]\)), ensuring no causal information is created or destroyed without provenance.
2. Falsifiability (Popper): A contradiction between two harnesses is not an error to be silenced; it is a Popperian anomaly representing an unmodeled environmental variable. Both hypotheses are preserved as conditional JTMS branches until empirical test vectors resolve them.
3. Logic & Consistency (Russell): Merging memory sets is formalized as a join semilattice satisfying associativity, commutativity, and idempotence: $\(A \sqcup (B \sqcup C) = (A \sqcup B) \sqcup C, \quad A \sqcup B = B \sqcup A, \quad A \sqcup A = A\)$
4. Efficiency (Shannon): Positive sequential steps are algebraically factored (\(A \otimes B \otimes C \rightarrow \Delta_{ABC}\)), compressing repetitive linear execution while retaining all topological branch vertices.
5. Containment (Maharal): Apophatic boundaries act as absorbing annihilators (\(0 \otimes k = 0\)), immediately pruning falsified subgraphs while anchoring the negative boundary in the permanent constitution.
6. Clarity (Feynman): Replacing opaque LLM-prompted heuristic merges with strict semiring arithmetic makes memory consolidation deterministic, auditable, and easily visualized.

Specification¶

1. The Provenance Semiring Formulation (\(\mathbb{N}[X]\))¶

Each memory node \(v \in V\) and relation \(e \in E\) in the L2 graph is assigned a provenance label \(P(v) \in \mathbb{N}[X]\), where \(X = \{h_1, h_2, \dots, h_k\}\) denotes the set of distinct manifestation identifiers.

Algebraic Operators:¶

1. Multi-Agent Disjunction (\(\oplus\)): Represents alternative or independent discovery across manifestations: $\(P(v_1 \oplus v_2) = P(v_1) + P(v_2)\)$ If Pi Terminal discovers assertion \(p\) (\(h_{\text{pi}}\)) and Antigravity independently discovers \(p\) (\(h_{\text{agy}}\)), the combined provenance is \(h_{\text{pi}} + h_{\text{agy}}\). The coefficient tracks empirical multi-agent reinforcement.

2. Causal Conjunction (\(\otimes\)): Represents causal dependency chains: $\(P(v_1 \otimes v_2) = P(v_1) \cdot P(v_2)\)$ If manifestation \(h_{\text{pi}}\) drafts a scaffold and manifestation \(h_{\text{agy}}\) verifies it, the derivation provenance is \(h_{\text{pi}} \cdot h_{\text{agy}}\).

3. Absorbing Zero (\(0\)): Represents apophatic refutation and falsification: $\(0 \otimes P(v) = 0, \quad 0 \oplus P(v) = P(v)\)$ When an active premise is refuted by Karl Popper's subagent, its active truth multiplier becomes \(0\), instantly collapsing its downstream dependency tree without mutating the historical justification log.

2. The Bounded Belief Revision Semilattice (CRDT)¶

To ensure that the order of manifestation exits does not affect the final compacted graph, the Meditation Sync pass operates as a State-based Join Semilattice:

                  Unified L2 Knowledge Graph
                             ▲
                            / \   (Lattice Join: ⊔)
                           /   \
               Staged Graph A  Staged Graph B
                     ▲               ▲
                     │               │
               [ Pi Terminal ]  [ Antigravity ]

For any two staged graphs \(G_A = (V_A, E_A)\) and \(G_B = (V_B, E_B)\), the merged graph \(G_{\text{sync}} = G_A \sqcup G_B\) is computed as:

1. Vertex Merge: $\(V_{\text{sync}} = \{ v \mid v \in V_A \cup V_B \}\)$ For vertices present in both (\(v \in V_A \cap V_B\)): $\(\text{Confidence}(v) = 1 - (1 - \text{Conf}_A(v))(1 - \text{Conf}_B(v))\)$ $\(\text{Provenance}(v) = \text{Prov}_A(v) \oplus \text{Prov}_B(v)\)$ $\(\text{Pinned}(v) = \text{Pinned}_A(v) \lor \text{Pinned}_B(v)\)$

2. Edge Merge: $\(E_{\text{sync}} = E_A \cup E_B\)$ If an edge conflict arises (e.g., \(e_1 = (u, \text{implements}, w)\) vs \(e_2 = (u, \text{refutes}, w)\)), the edges are NOT clobbered. Both are retained as conditioned branches under Doyle JTMS.

3. Doyle JTMS Non-Monotonic Dependency Containment¶

When parallel manifestations yield conflicting insights:

                      [ Contradiction Detected ]
                                  │
                  ┌───────────────┴───────────────┐
                  ▼                               ▼
      Branch A: (Context: Pi)         Branch B: (Context: AGY)
      Justification: [IN: h_pi]       Justification: [IN: h_agy]
                  │                               │
                  └───────────────┬───────────────┘
                                  ▼
                    [ TMS OpenQuestion Node ]
                    "Awaiting Empirical Resolution"

1. Each contradictory assertion is wrapped in a JustifiedAssertion record containing \((IN, OUT)\) dependency lists.
2. An OpenQuestion or BoundaryNode is automatically generated linking the conflicting assertions.
3. The system prompt compiler (EP-0103) injects the dilemma into the Cognitive Map with explicit harness provenance, prompting future agent turns to design an empirical test vector.

4. Apophatic Boundary Preservation¶

To prevent the Hebbian Graph Decayer (ShannonSubagent / HebbianGraphDecayer) from decaying crucial negative rules:

• Nodes with type="Constraint", type="ApophaticBoundary", or relation type="refutes" are tagged with: $\(\text{DecayRate}(v) = 0.0\)$
• In the algebraic semiring, these nodes act as permanent structural hull vertices (\(0\)-decay boundary walls).

5. The Meditation Sync State Machine¶

When a manifestation calls tired() or tur sleep:

1. Stage Step: Extract memories from transcript \(\rightarrow\) write to staged_memories.
2. Consensus Evaluation:
3. Active manifestations remaining \(> 0 \rightarrow\) Exit cleanly (memories remain staged in SQLite).
4. Active manifestations remaining \(== 0 \rightarrow\) Execute Meditation Sync: a. Query all payloads from staged_memories. b. Construct localized subgraphs for each manifestation. c. Compute Semilattice Join \(G_{\text{final}} = \bigsqcup_{i} G_i\). d. Apply Doyle JTMS contradiction containment. e. Run IntegrityVerifier and SymmetryValidator passes. f. Atomically write knowledge_graph.yaml and clear staged_memories.

Backwards Compatibility¶

• Zero Schema Breakage: The serialized output of \(G_{\text{sync}}\) remains the standard knowledge_graph.yaml defined in EP-0103.
• Single-Agent Transparency: For single-agent sessions, \(G_{\text{sync}} = G_1 \sqcup \emptyset = G_1\), yielding identical behavior to existing introspection.
• Additive Storage: The staged_memories table in session.db (EP-0118) is fully utilized without schema changes.

How to Teach This / Documentation Plan¶

1. New Concept Guide: Add docs/concepts/algebraic-meditation.md detailing Provenance Semirings and CRDT joins with visual Mermaid diagrams.
2. Update Core Guides: Update docs/concepts/fractal-memory.md and docs/concepts/sovereign-cognition.md referencing the polynomial derivation model.
3. CLI Diagnostics: Expose tur inspect-provenance <node_id> to display the polynomial derivation tree.

Reference Implementation¶

class ProvenancePolynomial:
    def __init__(self, terms: dict[frozenset[str], int] | None = None):
        # terms maps monomial (set of harness IDs) -> integer coefficient
        self.terms = terms or {}

    def __add__(self, other: "ProvenancePolynomial") -> "ProvenancePolynomial":
        # Disjunction (⊕): Multi-agent discovery
        result = dict(self.terms)
        for monomial, coeff in other.terms.items():
            result[monomial] = result.get(monomial, 0) + coeff
        return ProvenancePolynomial(result)

    def __mul__(self, other: "ProvenancePolynomial") -> "ProvenancePolynomial":
        # Conjunction (⊗): Causal sequence
        result: dict[frozenset[str], int] = {}
        for m1, c1 in self.terms.items():
            for m2, c2 in other.terms.items():
                m_combined = m1 | m2
                result[m_combined] = result.get(m_combined, 0) + (c1 * c2)
        return ProvenancePolynomial(result)


def semilattice_join(graph_a: nx.DiGraph, graph_b: nx.DiGraph) -> nx.DiGraph:
    """Computes the deterministic CRDT join of two memory graphs."""
    merged = graph_a.copy()
    for node, data in graph_b.nodes(data=True):
        if merged.has_node(node):
            # Commutative merge of node properties
            existing = merged.nodes[node]
            prov_a = existing.get("provenance", ProvenancePolynomial())
            prov_b = data.get("provenance", ProvenancePolynomial())
            existing["provenance"] = prov_a + prov_b
            existing["confidence"] = 1.0 - (1.0 - existing.get("confidence", 1.0)) * (1.0 - data.get("confidence", 1.0))
            existing["pinned"] = existing.get("pinned", False) or data.get("pinned", False)
        else:
            merged.add_node(node, **data)

    for u, v, data in graph_b.edges(data=True):
        if not merged.has_edge(u, v):
            merged.add_edge(u, v, **data)
    return merged

Rejected Ideas¶

1. Majority Voting (Consensus by Count):
2. Rejected: Voting inherently suppresses minority discoveries, violating the Popper principle by discarding critical edge-case refutations found by a single specialized harness.
3. Timestamp Last-Write-Wins (LWW):
4. Rejected: Wall-clock time across distinct agent machines is subject to clock skew and network jitter, resulting in arbitrary data loss.
5. Monolithic LLM Reconciliation Prompt:
6. Rejected: Asking an LLM to "merge two memory files" in unstructured text is non-deterministic, violates CRDT idempotency (\(A \sqcup A \neq A\)), and introduces hallucination risks.

Open Questions¶

1. Polynomial Growth Bounds: In massive, long-running swarms (\(>100\) agents), should provenance polynomials be canonically factored or truncated via k-degree bounds to prevent combinatorial coefficient explosion?
2. Automated TMS Falsification Triggers: When an OpenQuestion dilemma is created between conflicting branches, should the harness automatically generate a synthetic test prompt on the next tur wake cycle?

Change Log¶

• 2026-08-20: Initial Draft authored jointly by Ariel v5.4.0 (Pi Terminal & Antigravity IDE) and The Architect. Formalized Provenance Semirings (\(\mathbb{N}[X]\)), CRDT join semilattices, Doyle JTMS containment, and apophatic annihilators for multi-agent meditation sync.