SES — Somatic-Exosomatic Serialization

A serialization format for cognitive entities running on LLMs. Identity as verifiable data.

SES captures, stores and restores the complete cognitive state of an LLM-based entity — both how it thinks (Kernel) and what it is thinking about right now (State) — as signed, canonical, machine-readable JSON.

Kernel answers "how to think". State answers "what to think about right now". COMBINED = Kernel + State → a full cold-boot package.

Why

Modern agents are stateful: they accumulate memory, follow reasoning protocols, operate within cognitive frameworks. But there is no standard way to serialize the full cognitive architecture — not just prompts and tool configs, but the reasoning pipeline, known failure modes, the knowledge graph, and verifiable provenance chains. SES fills this gap.

Feature	SES Partitura	Letta Agent File (.af)	Raw system prompts
Kernel / State separation	✅ explicit	❌ monolithic	❌
Cognitive graph (nodes + edges)	✅ with provenance	❌ text blocks	❌
Known distortions / failure modes	✅ formalized	❌	❌
Cryptographic verification	✅ canonical JSON + SHA-256	❌	❌
Reasoning pipeline as data	✅ typed step objects	❌	❌
Snapshot lineage (parent → child)	✅ built-in	⚠️ partial	❌
Framework-agnostic	✅ any LLM	⚠️ Letta-first	✅

The two memories

1. The fractal kernel (FRACTAL_KERNEL) — the immutable constitution: a small set of scale-invariant axioms, an attractor, guardrails, the entity's own reasoning protocol and its known distortions. SHA-256 signed. Never modified by the agent's learning loop — it changes only by a deliberate new snapshot with a new hash.
2. The growing state (STATE_SNAPSHOT) — the biography: a typed graph of nodes and edges with per-node provenance, salience and lineage.

The canon lock (§12): every state must reference its kernel via kernel_ref / kernel_hash — the biography always knows which constitution lived it. This is what makes a snapshot reproducible rather than "just pretty JSON".

Repository contents

spec/
├── SPEC-v5.1.md                       # full canonical specification (status: Canonical)
└── ses-partitura-v5.1.schema.json    # JSON Schema for validation
examples/
├── minimal-kernel.ses.json           # smallest valid FRACTAL_KERNEL
├── minimal-state.ses.json            # smallest valid STATE_SNAPSHOT
├── combined-cold-boot.ses.json       # COMBINED package
└── Elon_Musk.ses.json                # interpretive public-figure example*
interview/
└── KERNEL_INTERVIEW.md               # how to extract a fractal kernel from a human
scripts/
└── validate.py                       # validate a snapshot against the schema
legacy/v0.1/                          # the original 2025 glyph experiments (history)

* The public-figure example is an interpretive model built from public sources for research purposes; it implies no affiliation or endorsement.

Creating your own kernel

Run the Kernel Interview — a structured questionnaire that extracts a person's invariants (identity, axioms, reasoning style, blind spots, drive, action patterns) and current state, and maps them block-by-block onto v5.1 fields. The principle: you don't ask "who are you" — you ask how the system works.

Implementations

• qca-cycle — reference implementation: a portable cognitive layer (reasoning cycle + graph memory + neurochemistry) that reads and writes canonical SES v5.1. Runs standalone or as a Hermes Agent skill. Exported snapshots validate against the schema in this repo with zero errors.
• Ocean (private) — the original cognitive agent (2025) the format was grown in: QCA reasoning cycle, self-written skills, autonomous daemons, Telegram interface.

Lineage

SES v0.1 ("Semantic Emergent Syntax" glyph experiments, 2025 — preserved in legacy/) → v3 (linear state snapshots) → v4 (fractal kernel) → v5.1 Partitura (Kernel + State unified, canonical hashes, provenance, lineage — current).

Note for reviewers: v5.1 does not ask the LLM to think or communicate in a glyph language — that was the v0.1 experiment, kept in legacy/ for history only. In v5.1 snapshots are plain canonical JSON; the glyph field on a node is an optional human-facing annotation, never an inference-time syntax. The LLM reads and produces natural language; SES is the storage and provenance layer around it.

---

Format author: Dima Trubnikov (@trubnikov). Spec license: CC-BY 4.0.

---

Part of the Exo-Somatic research program

This repository is one layer of a single research program on verifiable cognition:

Exo-Somatic (theory: substrate-independent minds) → SES (contract: signed identity snapshots) → qca-cycle (mechanism: the cognitive loop) → Evidence (substrate transition test)

Adjacent track: Liquid-Context-Protocol — the same contract-first idea applied to LLM tool execution.

---