Skip to content

Illustrative design animation · no measured traffic or test results

In this conceptual animation, points move through three proposed checks. Some stop and some continue. Their movement illustrates the design, not the behavior or reliability of a deployed system.

EMEBFilter Duplicates
EARLCheck Quality
ECGLVerify Identity
Data Pipeline
→
→
→
Emily · Science

How Emily handles memory and trust.

Useful memory needs provenance, review and correction. Emily’s architecture is designed to separate received information from the authority to change behavior or take action. Explore the design, its evidence status and the limits of what has been demonstrated.

The problem

What useful memory needs.

Continuity depends on more than keeping records. Sources, corrections, relevant retrieval and permission boundaries all matter.

Repeated or conflicting information

Similar wording can repeat an existing idea or introduce an important correction. Memory needs to preserve the difference, including the source and version of each contribution.

Errors carried forward

A saved statement can still be wrong. Confidence, fluent wording and successful storage do not establish truth; later answers need relevant sources and review.

Data mistaken for authority

Remembering a contributor’s statement should not grant that contributor permission to change identity, policy or behavior. Those boundaries need explicit controls and tests.

Three protocols

Design, implementation and evidence.

These protocol descriptions explain design intentions. Implementation and evaluation must be checked against a named release; no versioned deployment evidence is attached to this page yet.

Design stage 1 · The filter

EMEB

Expected Memory Error Bounds

EMEB is a design for assessing whether incoming information overlaps with existing memory. The aim is to reduce unnecessary repetition while preserving useful differences.

Design:
Described here; paper linked below
Implemented:
Deployment evidence pending
Evaluated:
Versioned test results pending
Version:
Paper and deployed versions not verified
Content reviewed:
16 October 2026

Design overview

The proposed approach uses semantic representations and probabilistic estimates to compare candidate memories. Similarity is a signal for review; it does not prove two statements are equivalent.

Limitations: Paraphrases, changed facts and conflicting sources need separate tests. Duplicate suppression is not a guarantee that an idea is stored only once, or that a correction is preserved.

Read the EMEB paper (opens in a new tab)

Design stage 2 · The quality check

EARL

Episode-level Adaptive Risk Learning

EARL explores how risk signals could guide the review of information before it is relied on. A confident answer can still be wrong.

Design:
Described here; paper linked below
Implemented:
Deployment evidence pending
Evaluated:
Versioned test results pending
Version:
Paper and deployed versions not verified
Content reviewed:
16 October 2026

Design overview

The design considers uncertainty and consistency signals. Which signals can be measured depends on the provider, model and interface. Closed-model APIs must not be assumed to expose attention, internal states or token-level probabilities.

Limitations: No provider-by-provider signal inventory or evaluated thresholds are attached here. This page does not claim that unreliable thoughts are always detected or prevented from being stored.

Read the EARL paper (opens in a new tab)

Design stage 3 · The gatekeeper

ECGL

Epistemic Confidence-Gated Learning

ECGL describes a boundary between receiving information and allowing it to influence identity or behavior. An attributed record can remain unverified DATA.

Design:
Described here; paper linked below
Implemented:
Deployment evidence pending
Evaluated:
Versioned test results pending
Version:
Paper and deployed versions not verified
Content reviewed:
16 October 2026

Design overview

The design separates the record of a contribution from the authority to change protected state. Any implemented gate needs explicit authorization, review and auditable decisions; a confidence score alone is not evidence of truth or permission.

Limitations: Deployed identity-change controls, owner/clone isolation and revocation require versioned tests. Owner-approved, scoped teaching grants are still in development.

Read the ECGL paper (opens in a new tab)
Together

The Epistemic Firewall.

The Epistemic Firewall names the proposed combination of these controls. The static explanation below also describes the animation: overlap, risk and authority are separate questions. A record being saved does not mean it is verified, retrieved correctly or authorized to change behavior.

  1. In

    Raw conversation

    Received information

  2. EMEB

    Assess overlap

  3. EARL

    Assess risk

  4. ECGL

    Review authority

  5. Memory

    Source and review state

Conceptual design: assess overlap, assess risk and review authority. This diagram does not establish that every deployed write follows this sequence or passes each check.

From a contribution to an answer

Connected assistants can contribute attributed information. An unverified DATA record preserves a contribution without establishing it as confirmed knowledge. Evidence of receipt, durable storage, retrieval, correct answer use and any resulting action must be checked separately.

Recall reliability and owner-controlled teaching permissions are in development. A fresh conversation may miss stored or selected context; review important answers against their sources. A failed or missing write is not a learned fact.

Inspect the contribution example
For developers

Build with Emily.

Explore the documented MCP connection or request access to the invite-only SDK. Client compatibility and reliable cross-session recall require separate tests for the versions you use.

EmilyRemote SDK · Invite-only

SDK and package access is currently by invitation. Contact us about access and the current package and documentation.

Request SDK access

MCP connection guide

Connect Emily to compatible AI tools.

Open the MCP connection guide

Developer Playground

Benefits, ways in, custom interfaces and how to get access.

Open the Developer Playground

Talk to us about Emily.

Request a technical briefing on how Emily's memory works, or a demonstration of a workflow under your brand.

Or email sales@robsoninc.com