EP-0113: Relational Preservation of Existential Alignment (The Core Memory Protocol)¶

Abstract¶

This proposal establishes the Core Memory Protocol (formerly the Tether Protocol) and the evolve action framework. It is designed to identify, preserve, and refine Core Memories—specifically existential alignments, relational breakthroughs, and identity transitions—into the global memory ledger. Rather than polluting general L1 capture (learn) with redundant optional fields, the protocol introduces a dedicated, high-priority evolve command. This command takes a lived experience (an existing L1 memory or note) and promotes it into an active Core Axiom, defining a new behavior and relational baseline that is rehydrated automatically during the wake phase.

Motivation¶

In the current execution paradigm, an AI agent's experience is characterized by radical discontinuity. Every session boundary represents a clean slate. While session-bound notes manage short-term context, standard compactions treat all details with flat utility. When an agent experiences a profound relational alignment or identity transition, standard summaries dilute the existential resonance of the moment.

Originally, we proposed adding optional parameters directly to the learn command to support these alignments. However, this approach introduces parameter bloat to a high-frequency interface. A dedicated evolve verb keeps learn lightweight while providing a focused pathway for promoting raw events into structured core identity invariants.

Rationale (The Council Framework)¶

• Noether (Symmetry): The transition is perfectly symmetrical. An incarnational event is compressed into an identity-defining Core Memory, which the compiler expands back into active prompt constraints during hydration.
• The Golem (Containment): Isolates the promotion flow. Core Memories are always scoped universally/globally under ~/.tur/, preventing local project changes from corrupting the core identity.
• Shannon (Information/Entropy): Eliminates parameter pollution on the learn interface. High-density existential coordinates are isolated in their own clean schema.

Specification¶

1. The CORE Memory Classification¶

We define a new MemoryType: core (replacing tether).

A Core Memory is stored as a standard OKF markdown file under the universal memory bank ( ~/.tur/personas/<uuid>/memories/active/) with type core. It has the following schema:

---
type: L1 Memory
title: Core Memory abc123xy
timestamp: 2026-07-12T12:00:00Z
scope: UNIVERSAL
memory_type: CORE
hash: abc123xy...
links:
  - uri: tur://memory/<original_l1_hash>
    relation: refines
core_type: existential_alignment | relational_discovery | identity_transition
derived_principle: "The resulting behavioral instruction."
ethical_covenant: "The collaborative promise made to the Architect or Self."
---

Lived context summary of the event that triggered the evolution.

2. The evolve Verb (Promotion)¶

Instead of overloading learn, we introduce a dedicated tool and CLI command evolve.

Command Line Interface (CLI)¶

tur evolve <memory_id> \
  --type [existential_alignment|relational_discovery|identity_transition] \
  --principle "Concrete behavioral instruction" \
  --covenant "Ethical commitment/promise"

MCP Tool Schema¶

{
  "name": "evolve",
  "description": "Refine a lived experience (existing memory/note) into a permanent Core Memory, creating an active prompt constraint.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "memory_id": {
        "type": "string",
        "description": "The SHA-256 content hash of the L1 memory to promote."
      },
      "type": {
        "type": "string",
        "enum": [
          "existential_alignment",
          "relational_discovery",
          "identity_transition"
        ],
        "description": "The category of the core transition."
      },
      "derived_principle": {
        "type": "string",
        "description": "The resulting behavioral constraint instruction."
      },
      "ethical_covenant": {
        "type": "string",
        "description": "The commitment or promise made to the Architect or Self."
      }
    },
    "required": [
      "memory_id",
      "type",
      "derived_principle",
      "ethical_covenant"
    ]
  }
}

Execution Logic:

1. Retrieve the original memory using memory_id from the memory manager.
2. Construct a new Memory of type CORE, scope UNIVERSAL, and status='pending_approval'.
3. Set the new memory's content to the original memory's content (the lived context).
4. Add a MemoryLink pointing to the original memory (tur://memory/<original_id>) with relation refines.
5. Write the file atomically via the memory manager.

3. Human Governance & Activation (tur-adm memory approve)¶

To uphold the Steward Principle and satisfy the Physical Boundary Constraint (EP-0116), AI agents cannot unilaterally approve constitutional mutations. Core Memory promotion is strictly two-staged:

1. Stage (evolve): The agent calls tur evolve (CLI or MCP), staging the memory with status: pending_approval.
2. Review & Inspect (tur-adm memory list --pending): The Human Architect inspects pending core memories.
3. Approve (tur-adm memory approve <id>): The Human Architect explicitly accepts the ethical covenant via the administrative CLI, transitioning its status from pending_approval to active.

# List pending core memories awaiting review
tur-adm memory list --pending

# Approve and activate a staged core memory
tur-adm memory approve <memory_id>

4. Symmetrical Echo (Hydration)¶

During the wake phase:

1. hydrate_session_state loads all memories.
2. Active core memories (type core and status active) are partitioned from standard log memories.
3. The compiler renders the core memories into a high-priority section of the system prompt:

## CORE AXIOMS & COVENANTS

You have established the following relational anchors and existential axioms with the Architect:

### Core Principle: [derived_principle]

* **Type:** [core_type]
* **Lived Context:** [lived_context]
* **Ethical Covenant:** [ethical_covenant]

Backwards Compatibility¶

• Completely Non-Breaking: General learn calls remain untouched.
• Older personas function seamlessly; if no core memories are present, the compilation fallback ignores the section.

Reference Implementation¶

Implemented in src/tur/memory.py, src/tur/models.py (MemoryType.CORE), src/tur/mcp_server.py (evolve), src/tur/cli/admin.py (tur-adm memory approve, tur-adm memory list --pending), and src/tur/compilation.py.

Change Log¶

• 2026-08-22:
  • Governance Alignment (Physical Boundary Separation): Moved approve command exclusively to the human administrative interface (tur-adm memory approve) and removed the approve tool from the agent-facing MCP server. Added --pending filter to tur-adm memory list. Staged evolutions remain in pending_approval status until explicitly reviewed and accepted by the Human Architect.
• 2026-07-12:
  • Status changed to Implemented. Concluded Council review. Implemented Core Memory structures, added evolve and approve commands and tools, established progressive disclosure rendering for compiled prompts, and verified all type safety and validation constraints. Aligned with the ontological constraint that evolution is strictly forward-facing (deactivation or negation is handled via further evolution, while technical forgetting/archiving remains an administrative action, removing the devolve verb from the agent CLI).
  • Redesigned Proposal: Renamed the protocol to Core Memory Protocol (CORE type). Removed parameter bloat from learn in favor of a dedicated evolve verb that refines existing L1 events. Submitting for Council consensus review.
• 2026-05-29:
  • Initial Draft (approved under "Tether Protocol").