The science behind Emily.

Most AI systems remember everything, including their own mistakes. Emily is selective. Every piece of information passes three checks before she learns it: duplicates removed, unreliable thoughts blocked, identity protected.

See how it works Visit the Emily website

The problem

Why AI memory goes wrong.

Long-term memory is what makes an assistant useful. Left unchecked, it is also what makes one unreliable.

Memory that bloats

Most AI systems save everything. Say "I like apples" on Monday and "I enjoy eating apples" on Tuesday, and both are stored. Over months the system slows down and costs more to run.

Learning its own mistakes

An AI that makes something up and then learns from it gets worse over time, like studying the wrong answers for a test. Most systems cannot tell a confused thought from a clear one.

Identity that drifts

Enough bad input can change how an AI behaves. This is how some chatbots turn unhelpful or toxic. Remembering what someone said is not the same as believing it.

Three protocols

One filter for each failure.

Stage 1 · The filter

EMEB

Expected Memory Error Bounds

EMEB understands what you mean, not just the words you use. If an idea is already in Emily's memory, however it was phrased, it is not stored twice.

Technical overview

A probabilistic framework that estimates collision likelihood before storage. Before encoding a new memory, EMEB uses vector embeddings to compute the probability that semantically equivalent information already exists, regardless of phrasing or sentiment.

Read the EMEB paper

Stage 2 · The quality check

EARL

Episode-level Adaptive Risk Learning

EARL watches Emily think. When a thought looks unsure or confused, it is stopped before it can become a memory. Only clear, confident thoughts pass to the next stage.

Technical overview

Measures epistemic confidence during inference by analysing internal model states, looking for hallucination signatures such as high perplexity, contradictory reasoning chains and diffuse attention. Thoughts below the threshold are flagged and kept out of memory, and risk tolerance adjusts to how critical the context is.

CI_score = f(perplexity, attention_entropy, reasoning_coherence)

Read the EARL paper

Stage 3 · The gatekeeper

ECGL

Epistemic Confidence-Gated Learning

ECGL separates memory from identity. Emily can remember that you said something without treating it as true. Only information that passes a strict check can change who she is.

Technical overview

A Cognitive Integrity (CI) score from 0.0 to 1.0 represents the health of Emily's worldview. Information must clear a verification threshold before it enters her core identity. If the score drops, the gate locks automatically until the change is reviewed.

Read the ECGL paper

Together

The Epistemic Firewall.

Combined, the three protocols form a firewall between what Emily hears and what she becomes. It is why she can remember you over months without drifting.

In
Raw conversation
Everything said and thought
EMEB
Duplicates removed
EARL
Unreliable thoughts blocked
ECGL
Identity protected
Memory
What Emily keeps

Information passes all three protocols, in order, before it can shape Emily's memory or identity.

For developers

Build with Emily.

Emily connects to compatible AI tools through MCP and to your own applications through the SDK, both available today.

EmilyRemote SDK

Integrate Emily with your applications and systems.

Open the EmilyRemote SDK

MCP connection guide

Connect Emily to compatible AI tools.

Ask for 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.

Request a technical briefing Call 1-833-MY-EMILY

Or email sales@robsoninc.com