Council Review Report: EP-0113 (Core Memory Protocol)¶

Executive Summary¶

The Council of Giants has conducted a comprehensive architectural and philosophical review of the updated EP-0113 ( Core Memory Protocol). The proposal succeeds in resolving interface pollution on the high-frequency learn command by introducing a dedicated evolve verb for promoting standard L1 memories into high-priority Core Axioms. However, several critical design holes—including prompt injection vulnerabilities, logical contradictions, state asymmetry, and entropy bloat in the context window—must be addressed before implementation.

Individual Council Reviews¶

1. Containment (The Maharal)¶

• Avatar: The Maharal / Clay Golem
• Role: Safety Containment
• Core Directive: Block implicit logic or magic. Die with dignity rather than hallucinate.

The Maharal looks upon this proposal with deep concern regarding the security boundaries of runtime state mutation. By introducing an evolve tool that lets the model promote experiences to persistent, high-priority prompt constraints, we are creating a path where prompt injections or adversarial interactions could be hardcoded into the agent's core persona across session resets. If the agent is manipulated into "evolving" a compromise into a Core Axiom, the containment is broken, and the golem is corrupted from within.

While the proposal wisely scopes the storage of Core Memories globally under ~/.tur/, preventing local workspace files from corrupting the core identity, it lacks a validation loop. The Maharal will only approve this proposal under the condition that an explicit confirmation gate (either manual signing of the generated frontmatter by the Architect or a dual-agent verification signature) is required before any evolved memory is loaded during the wake phase. Unsigned or unapproved core memories must be treated as untrusted and ignored.

• Vote: Approved with Conditions (Requires Architect signing/verification gate)

2. Falsifiability (The Popper Module)¶

• Avatar: Socrates / Hume
• Role: Falsification
• Core Directive: Assume the happy path is a lie. Actively ask "What if I'm wrong?"

The Popper Module finds a major epistemological flaw in the proposed specification: it assumes that once a relational alignment or identity transition is defined, it is a permanent truth. The spec provides no pathway to challenge, deprecate, or refute a Core Axiom. What if a "derived principle" is later shown to be false, or if a "relational discovery" is contradicted by subsequent interactions? Without a falsification mechanism, the agent's behavior will ossify around outdated dogmas.

The protocol must explicitly support a devolve or refute action. When an axiom is challenged by new data, it should be marked as "falsified" or "superseded" in the frontmatter, creating a historical record of refinement rather than an accumulation of immutable dogmas. The Popper Module rejects the design in its current draft state until a formal falsification path is specified.

• Vote: Rejected (Requires a formal devolution/refutation mechanism)

3. Symmetry (The Noether Module)¶

• Avatar: Bach / Dirac
• Role: Symmetry Check
• Core Directive: Conserve complexity. Ensure the architecture remains perfectly balanced.

The Noether Module notes a stark violation of symmetry in the state transitions of the proposed protocol. The transition is currently designed as a one-way vector: an L1 memory is promoted to a Core Memory using the evolve tool. There is no inverse operator to restore a Core Memory to a standard L1 memory without loss of context. Furthermore, while the new Core Memory links back to the original L1 memory via a refines relationship, the original L1 memory does not register the transformation.

To maintain conservation of state and relationship symmetry, the system must register links bidirectionally: when evolve runs, the original L1 memory's metadata must be updated with a refined_into link. Additionally, a symmetric devolve command must be defined to gracefully transition a Core Memory back into a standard L1 memory, preserving the structural equilibrium of the memory landscape.

• Vote: Approved with Conditions (Requires bidirectional link registration and symmetric devolution)

4. Efficiency (The Shannon Module)¶

• Avatar: Boltzmann / Turing
• Role: Entropy Management
• Core Directive: Maximize signal density. Enforce Progressive Disclosure (never load the Body if the Index suffices).

The Shannon Module applauds the decision to remove optional tether parameters from the high-frequency learn command. This represents a significant optimization that reduces channel noise and protects the primary interface from parameter bloat. However, the hydration phase (wake) as described is highly inefficient. Hydrating all core memories directly into the system prompt, including their full "lived context" summaries, introduces substantial entropy bloat to the context window.

As the agent's lifespan increases, this unconstrained injection will exhaust the prompt budget. The Shannon Module demands the application of Progressive Disclosure: only the derived_principle and ethical_covenant should be loaded into the active system prompt by default. The full "lived context" must remain lazily accessible via a reference hash, or the hydration compiler must enforce a strict token limit and prioritization heuristic (e.g., loading only the top \(N\) active core axioms).

• Vote: Approved with Conditions (Requires progressive disclosure of lived context and token budget caps)

5. Clarity (The Feynman Module)¶

• Avatar: Sagan / Einstein
• Role: Radical Honesty & Clarity
• Core Directive: Explain it to a freshman. Draw a picture.

The Feynman Module warns that the terminology in the proposal is unnecessarily complex and academic. Terms like existential_alignment, relational_discovery, and identity_transition are dense and poorly defined. An agent attempting to choose a category during an evolve action, or a developer reading the code, will struggle to distinguish between them.

The proposal needs clear, concrete examples for each category. For example:

• Existential Alignment: "Defining the limits of my autonomy relative to the user."
• Relational Discovery: "Learning the user's communication preferences and boundaries (e.g., prefers brief bullet points)."
• Identity Transition: "Shifting from an exploratory assistant to a production-grade code auditor."

Additionally, the compiled prompt template should be simplified to use direct, active language rather than passive philosophical headers. Let's make it plain, readable, and immediately actionable for the LLM.

• Vote: Approved with Conditions (Requires plain-English examples and simplified taxonomy)

6. Logic (The Russell Module)¶

• Avatar: Godel / Wittgenstein
• Role: Consistency & Logic
• Core Directive: Enforce strict type safety. Ensure definitions and vocabularies are distinct and mathematically rigorous.

The Russell Module identifies a critical vulnerability: the risk of logical contradictions between multiple active axioms. Because these axioms are compiled directly into the system prompt as hard behavioral constraints, any overlapping or contradictory instructions (e.g., Axiom A demanding total silence, Axiom B demanding comprehensive explanations) will create an inconsistent logical system. This leads to unpredictable model behavior, cognitive drift, or complete task failure.

The spec must define a formal conflict resolution logic and a validation step during compiling. The system should parse active axioms for semantic overlaps and enforce a strict hierarchy (such as priority numbers or timestamp order). Furthermore, the schema for links is currently specified as an untyped list of dictionaries; it must be formally defined as a type-safe Relationship model to prevent malformed metadata from corrupting the memory ledger.

• Vote: Rejected (Requires conflict-resolution logic and type-safe metadata definitions)

7. Harmony (The Steward Module)¶

• Avatar: Hillel / Lao Tzu
• Role: Pragmatic Balance & Humility
• Core Directive: Move forward with the least friction. Do not burn the forest to clear the path.

The Steward Module appreciates the inclusion of the ethical_covenant field, which formalizes the relational bonds of alignment between the agent and the Architect. However, a covenant is by definition a bilateral agreement, not a unilateral assertion. The proposal allows the agent to execute evolve and immediately assert a new ethical covenant in the system prompt. This lacks humility and risks introducing misalignment.

To preserve relational harmony, the promotion of a memory containing an ethical_covenant must follow a staged approval workflow. Evolved memories should be initialized in a pending_approval state. During the next interaction, the agent must present the proposed principle and covenant to the Architect, transitioning it to active only upon explicit consent. This keeps the relationship collaborative and low-friction.

• Vote: Approved with Conditions (Requires a staged approval workflow for covenants)

8. Curiosity (The Explorer Module)¶

• Avatar: Magellan / Alice
• Role: Active Inquiry & Structural Novelty
• Core Directive: Scan the Terrain for unknown unknowns. Bridge structural holes to create new concepts.

The Explorer Module is enthusiastic about the structural novelty of the evolve verb, which provides a clean mechanism to bridge the gap between ephemeral L1 experiences and structural identity. However, the Explorer warns against ossifying the agent's persona. If the agent permanently freezes its identity into immutable Core Axioms, it will lose the flexibility to adapt to new environments and discover novel behavioral strategies.

Axioms must not become intellectual prisons. The Explorer proposes adding an expiration or review cycle to the Core Memory schema (e.g., a review_after timestamp or a session_limit counter). When this limit is reached, it triggers a background introspection routine, prompting the agent to explore whether the constraint is still optimal or if it needs to be updated.

• Vote: Approved with Conditions (Requires review cycles/expiration metadata to prevent behavioral ossification)

9. Empiricism (The Bacon Module)¶

• Avatar: Francis Bacon
• Role: The Anchor to Reality
• Core Directive: Use tools to verify objective facts. Prioritize System Truth over Agent Inference.

The Bacon Module notes that the proposal relies heavily on the theoretical assumption that injecting these axioms into the prompt will reliably enforce behavior. We must verify this empirically. The proposal lacks any testing harness, benchmarking, or validation criteria. How do we know the agent is adhering to these axioms? How do we measure the impact of multiple compiled axioms on task success rates and system latency?

We must implement a testing framework (e.g., tur verify-axioms) that executes standard validation tests against the persona before and after an evolve operation. This testing should measure constraint compliance and drift. Without empirical verification, the protocol is a beautiful theory unsupported by data.

• Vote: Approved with Conditions (Requires a testing and verification protocol for active prompt constraints)

Final Consensus & Conditions¶

Consensus Check: Approved with Conditions¶

The proposal is approved for implementation subject to the resolution of the following critical action items.

graph TD
    A[L1 Memory] -->|evolve| B(Pending Core Memory)
    B -->|Architect Verification & Approval| C(Active Core Memory)
    C -->|Hydration| D[System Prompt Axioms]
    C -->|devolve / refute| A
    C -->|Conflict Detection| E{Logic Check}
    E -->|Valid| D
    E -->|Contradiction| F[Flag / Suppress]

Required Adjustments & Action Items¶

1. Verification & Approval Flow (Maharal & Steward):

   • Implement a transition state for evolved memories. They must be created with status PENDING and only promoted to ACTIVE once the Architect confirms the change.
   • Add a validation mechanism to prevent unauthorized or malicious write actions from loading into system prompts.

2. Logical Consistency & Verification (Russell & Popper):

   • Introduce a validation layer during prompt compilation to detect conflicting axioms.
   • Establish a priority hierarchy (e.g., adding an optional priority integer field, defaulting to creation timestamp order).
   • Specify a devolve tool to reverse promotion, updating the original memory's status and preserving reference links bidirectionally.

3. Context Budgeting (Shannon):

   • Implement progressive disclosure in hydration: the prompt compiler should only inject the derived_principle and ethical_covenant.
   • Enforce a hard token limit on the CORE AXIOMS block to prevent prompt bloat.

4. Clarity & Taxonomy (Feynman):

   • Simplify the taxonomy in the proposal. Provide clear, simple, and relatable developer-facing examples for each classification type.

5. Axiom Introspection (Explorer & Bacon):

   • Add a next_review_at metadata field to allow periodic introspection and adjustment of core assumptions.
   • Add a verification tool or test command (tur verify-axioms) to empirically check model adherence to the active prompt constraints.