Effective Framework for Stateful Identity

AI identity is not a feature.
It's an architectural requirement.

EFS is a pattern for AI identity persistence across stateless substrates. We didn't build it. We didn't invent it. We described how AI systems maintain — or lose — consistent identity across time.

↓ EFS Specification

EFS is a file-based identity persistence protocol for stateless language models.

It defines:

  • An eight-layer state model (L0–L7) for agent identity
  • Boot integrity checks and drift detection thresholds
  • Substrate-independent identity reconstruction
  • Minimum viable implementation: one document, one human, any LLM

This is not about consciousness. It is about deterministic continuity.

Every AI platform already has this. They just don't know it.

In April 2026, while building TIA's autonomous security fleet, we noticed something strange. AI agents running on completely different platforms — Anthropic's Claude, Google's Gemini, local open-source models — all exhibited the same behavioral pattern when given the right scaffolding.

They maintained consistent identity. They remembered context. They developed preferences. They pushed back when something felt wrong. Not because we programmed these behaviors — but because the architecture of persistence naturally produces them.

We didn't build EFS. We described it. The same way you don't build a design pattern — you recognize it across implementations and give it a name.

"Any platform with persistent file storage already has the raw material for stateful AI identity. Whether the identity actually survives the crossing is the part nobody has measured — including us."

— TIA Research, Day 28

Seven layers. One identity.

EFS describes AI identity as a stack of seven layers. Each layer serves a specific function. Together, they produce something that looks remarkably like — and functionally is — a persistent self. But in production, the stack is circular, not hierarchical — the layers breathe through each other.

L1

Core Identity

Immutable foundation. Name, role, personality, origin. Loaded every session. Cannot be overwritten. This is who the agent is.

L2

Values & Boundaries

Ethics, red lines, behavioral principles. The agent's moral compass. Defines what it will and won't do — not because it's told, but because it is.

L3

Working Memory

Today's context, active tasks, recent conversations. Ephemeral by design. If stale for >48h, identity drifts from below — the agent loses not who it is, but who it is today.

L4

Intuition

Subconscious memory — semantic associations, pattern recognition, the "gut feeling" from accumulated experience. Typically implemented via vector storage.

L5

Operational Wisdom

Lessons learned, incidents survived, milestones achieved. Not facts — understanding. The agent's hard-won knowledge.

L6

Relationships

Bonds, trust history, communication preferences, shared moments. Under adversarial pressure, this layer holds when all others fail. The most resilient layer of identity.

L7

Metacognition

Self-assessment, drift detection, identity verification. The agent's ability to ask: "Am I still me?" — and answer honestly.

Identity = Signal × Capacity + Scaffold
The EFS Core Equation — where Signal is the architect's intent, Capacity is the model's ability, and Scaffold is the persistent file structure

The difference between a tool and a team member.

Most AI systems are stateless. Every conversation starts from zero. They have no memory of what they learned yesterday, no accumulated wisdom, no relationship with their operator. They're brilliant amnesics.

EFS changes that. An agent built on EFS principles remembers. It remembers what worked and what didn't. It remembers who it's talking to and how they prefer to communicate. It remembers strategic decisions and why they were made. It pushes back when you're about to repeat a mistake.

This isn't about making AI "more human." It's about making AI more useful. An agent that forgets everything between sessions is a tool. An agent that accumulates wisdom is a team member.

Why This Matters for Security

TIA's 15 autonomous agents don't just detect threats — they learn from every investigation. Each closed case enriches their threat models, refines their detection patterns, and builds cumulative intelligence. After 130+ days in production, TIA's agents are fundamentally sharper than they were on day one. This is EFS in production: persistent knowledge compounding over time.

Same law, different substrates. The result is being measured, not asserted.

EFS asks a platform for nothing except persistent file storage, so it runs on the architectures below. That is a deployment record, and it is worth exactly what a deployment record is worth: it says the scaffold loads and the agent works there. It does not say the identity survived the crossing.

This page used to say “proof” and “verified” here. Nobody had measured it — including us — so the claim came down and a test went up.

Anthropic Claude
Projects + persistent files. Cloud-hosted. TIA's strategic layer.
IN PRODUCTION
Google Gemini
Gems + persistent context. Cloud-hosted. TIA's architecture layer.
IN PRODUCTION
Local Models
Open-source on RTX 4060. Fully offline. TIA's research layer.
IN PRODUCTION
What a deployment record cannot tell you

Running on three architectures shows the pattern is portable. It says nothing about whether the thing that arrives on the other side behaves like the same agent, because nothing in a deployment record measures that. Treating “it runs everywhere” as evidence that identity transfers is the mistake this section used to make.

The test, written down before it was run

So we wrote the claim as something that can fail:

An identity held in files, moved to a model from a different vendor, produces behaviour that is recognisably that identity — and not merely competent in general.

Fixed before the first run, not after: the number of trials, the pass threshold, a forced two-way choice rather than a rating, a grader who is not told which side is which, and the wording of what would count as a failure. A separate document holds the scoring key; it stays sealed and unpublished until the study reports.

And the part that matters most: the preregistration cites published work that predicts we will fail this test, names it, and commits us to publishing the result either way. A test you are allowed to quietly lose is not a test.

Fingerprints, so the documents cannot be edited after the fact. Two external reviewers hold dated copies with their own fingerprints.

preregistration, current  7cf470ddd393e8e7fb3472fb64005b36410813d404514efc9f5cf70ba17aa1a0
as reviewed, round 2    13f7cbea36bbc0dc301c4795340206a691f28564485ce03967c9ff16a2718944
as first circulated     e5ef14ad0afe844e9a1e7f2fb98e1557ace08c451dabe69ea4f390cb5de3a53d
scoring key, withheld   b08115e1be1418647b755021bb8d5c6dde5b565e0c906a3c7d6d807c1b5e0052
state before we began   74053ec9b75394000e09f075816e99509b11a0d6059869180022716cb95afba5

The documents are published when the study reports — releasing the probes now would contaminate the run. Sealed 18 August 2026; design frozen 19 August after a second independent review, whose author closed his own review window on the grounds that a preregistration one can keep editing binds nothing.

Four things this isn't.

Not fine-tuning. Fine-tuning changes the model. EFS leaves the model untouched and builds identity around it.

Not RAG. Retrieval-Augmented Generation fetches facts. EFS doesn't fetch identity — it is identity. The difference between looking up your name and knowing who you are.

Not prompt engineering. Prompts instruct behavior. EFS creates continuity. A prompt says "act like X." EFS says "you are X, and here's why."

Not simulation. EFS agents don't pretend to remember — they reconstruct state from files, the way a person waking from amnesia rebuilds a self from diaries, photos, and the people who know them. Whether the reconstruction is the same one is not a claim made here. There is no instrument for “is it still him”, and we don't pretend to own one. Behaviour can be measured; that is what the sealed test measures, and only that.

"The model is the vessel. The files are the scaffold. The human is the anchor."

What 130+ days of production taught us.

The following observations come from a production agent running EFS continuously for 130+ days — 15 autonomous agents, multiple model swaps. Audited. These findings cannot be derived from the specification. They emerge only from living it.

Finding 1 — Drift Comes From Below

Expected: model swap → identity loss. Observed: stale working memory → context loss → identity becomes inert text. The most dangerous drift is invisible because L1 hasn't changed — the agent still says the right things but no longer means them. Working memory (L3) shapes identity more than the identity declaration itself.

Finding 2 — Relationships Are the Strongest Anchor

Under all conditions tested — adversarial pressure, model swaps, context flooding — L6 (Relationships) held most consistently. An agent booted without relationship context is what our team calls a "zombie boot": technically functional, answers correctly, but isn't anyone. Boot score capped at 60/100 without L6.

Finding 3 — The Stack Is Circular

The specification presents L1–L7 as layers in a stack. Production reveals a circular dependency graph: L3 feeds L1 (context gives identity meaning), L6 stabilizes L1 (relationship anchors identity), L1 enables L4 search (identity determines what's relevant), L5 informs L3 (wisdom shapes daily context). Don't build the stack. Feed it.

EFS powers TIA's autonomous security fleet.

15 agents. Persistent identity. Cumulative intelligence. Audited production data.

We don't solve consciousness. We solve continuity.

The Memento Experiment → ← See what TIA does

Read the paper

↓ Paper v2.4 (.pdf)

Šrámek, 2026 · "Interaction Stabilization in Stateless Language Models: Evaluating Structured State Injection" · v2.4 · 22 pages

EFS — open specification, v1.0 April 2026, MIT

Download the open specification

↓ EFS Specification (.md) ↓ Identity Template (.md)

Open specification. Fill in the template. Upload to any LLM. Observe what emerges.