BREW · Designing context you can see, hold, and reshape

01The Problem

Every conversation with an AI starts from zero.

You tell it who you are, what you need, what your constraints are. It holds that for exactly as long as the tab stays open. Close it, and the model of you is gone. But even within one conversation, the problem is worse than forgetting.

The AI does not just forget. It actively degrades.

As a conversation grows, the model loses track of what you said earlier. It makes premature assumptions before you have finished explaining. It anchors to its own earlier outputs and resists correction. And when it goes wrong, it often cannot recover, even when you tell it so directly.

The result is a fundamental mismatch. Users build up context across turns, but the AI's ability to use that context deteriorates as the conversation grows. The context is technically there, sitting in the chat history, but it is buried, unstructured, and increasingly invisible to the model. This is the within-session problem BREW is designed to fix.

39%

average performance drop measured across multi-turn conversations, the more you tell it, the more it loses

0

persistent, structured model of you that survives a closed tab or a mid-conversation pivot

100%

invisible, you cannot inspect, correct, or steer what the AI believes about you

The context is there. It is just buried, unstructured, and invisible to the very model that is supposed to use it.

02Background & Related Work

Standing where three research threads meet.

BREW sits at the intersection of long-context failure in language models, direct manipulation as an interaction philosophy, and prior attempts to reify memory and intent. We read widely before we built anything.

A

Long contexts fail

Language models use long contexts poorly, attention sags in the middle, and multi-turn performance degrades. The information is present but not usable.

B

Direct manipulation wins

Shneiderman's principle: people reason better about visible objects they can act on directly than about hidden state addressed through commands.

C

Memory is unsolved

Prior systems reify conversational memory and intent, but treat retrieval across sessions, not the live, steerable model of a person within one.

Reading list that shaped the work

Lost in the Middle: How Language Models Use Long Contexts

Evidence that models ignore information buried mid-context. Motivates externalizing context out of the transcript.

Long context

Direct Manipulation: A Step Beyond Programming Languages

Shneiderman's foundation: visible objects, reversible actions, immediate feedback. BREW's interaction spine.

Foundations

Direct Manipulation vs. Interface Agents

The Shneiderman and Maes debate. BREW argues you can have agency and an agent at once.

Foundations

DirectGPT: A Direct Manipulation Interface to Interact with LLMs

Bringing manipulable objects to LLM interaction. A close neighbor in the design space.

Manipulation + LLMs

Memolet: Reifying the Reuse of User-AI Conversational Memories

Reifies memory for cross-session reuse. BREW deliberately solves the orthogonal within-session problem.

Memory

Brickify: Expressive Design Intent Specification through Direct Manipulation

Blocks and bricks as a grammar for intent. Inspiration for our physical block game.

Manipulation

Don't Just Tell Me, Ask Me: AI Systems that Intelligently Frame

On framing and elicitation. Informs how BREW surfaces what it does not yet know.

Elicitation

Luminate & Boundary Objects

Structured exploration of design spaces, and artifacts that hold shared meaning between people and systems.

Structure

03Ideation

We made the abstraction physical before we made it digital.

Context architecture is abstract. To pressure-test the idea with real people, we built a hands-on block game: players construct their life context out of physical blocks, then weather a mid-conversation pivot and watch what breaks.

Blocks make the abstraction concrete. Stacking reveals weight and dependency. Moving a block shows ripple effects. Cloning a story across domains shows how one human moment can justify constraints in several places at once. Each session ran 8 to 12 minutes and produced both a physical structure and a set of narrative insights that fed directly into our research observations.

Scenario 01

The Thesis Crunch

How context collapses under pressure

A grad student with a Friday thesis, a half-marathon, a relationship, and an $80 budget. The advisor moves the deadline to Tuesday. The player rebuilds and watches a single hard constraint propagate across every domain.

Demonstrates
Cloning, lateral links12 min
3 rounds

Scenario 02

The Creative Block

How cloning builds new insight

A designer who suddenly cannot work. By cloning one story (a colleague left) across unexpected aspects, the player discovers the root cause is isolation, not skill. The clones make a hidden dependency visible.

Demonstrates
Cross-Part reasoning10 min
3 rounds

Scenario 03

The Budget Pivot

How context survives change

A product manager mid-sprint when the company pivots. The player shows which constraints survive when they are externalized and visible, versus which fall through the cracks when buried in chat and Slack.

Demonstrates
Constraint survival10 min
3 rounds

Who showed up at the exhibit

What we did not expect was who resonated. None of them were AI researchers. All of them recognized the frustration immediately.

MAn NYU math alum, class of the 1980s

He understood the problem before the second sentence was finished. What turned into a proper, crouched-down conversation showed the idea lands across generations of technical thinking.

LA lawyer

He saw it through the lens of disclosure and informed consent. If the AI holds a model of you, you have a right to see and correct it. Context becomes a question of rights.

AA senior product designer, Amazon

She started riffing on how this maps to enterprise AI workflows, where constraints span teams and a lost requirement is expensive. Context as shared, durable state.

04The Core Idea & Data Model

Context as a first-class object you live inside.

BLOOM externalizes the AI's working state. Instead of context buried in chat history where neither side can see it clearly, it becomes a visible, editable object that lives alongside the conversation. The chat handles dialogue. The context view handles the AI's model of you. They update each other, but stay structurally separate.

The context view is not a summary of the chat. It is a parallel representation of your world that the AI reads directly before generating every response. That makes BLOOM a soft explainability mechanism: it gives you enough visibility into the AI's working state to calibrate trust and intervene when it is wrong, without exposing model internals or demanding technical knowledge.

WWorld

The identity container. It tells the AI that everything below belongs to one person with one coherent life, so it can reason across domains: a knee injury, a thesis deadline, and an $80 budget are constraints on the same human, sometimes in tension.

PParts

The major domains of a life. Academic, Health, Finance, Relationships, Creative work. Each Part is a bounded region of the World, and the AI treats constraints from different Parts differently.

AAspects

The dimensions within a Part. Within Academic: Deadlines, Work schedule, Output, Energy. Aspects are the interpretive frame, they tell the AI what kind of thing a Point is. The Aspect is what gives a Point its meaning.

·Points

Machine-actionable facts. Concrete, specific, directly usable as constraints. "Thesis due Friday 11:59pm." "No shellfish, allergy." Everything below justifies a Point. Everything above contextualizes it.

Points are where meaning crystallizes.

"Stories

The raw human evidence. Your own words, unprocessed and emotional. The AI does not act on Stories directly. It reads them to understand why a Point exists and how heavily to weight it. The relationship is many-to-many: one Story can ground four Points across three Parts.

Interactive

Walk the hierarchy yourself

This is Maya's world, our running example. Click a Part to zoom into its Aspects, click an Aspect to see its Points, then open a Point to read the Stories that ground it. Zooming is the only navigation. The act of going deeper is the act of moving down the hierarchy.

Live demoA faithful recreation of BLOOM's semantic-zoom canvas, built from the project's real sample data.

05Constraint Types

Not all facts are equal. Every Point carries a certainty signal.

BLOOM tags every Point with one of three constraint types. They tell the AI how much flexibility it has, and they give it a decision hierarchy without asking users to assign numbers or understand model internals. Hard, conditional, and soft are vocabulary people already have.

Hard

Zero flexibility

The AI must never violate, suggest around, or treat these as negotiable. They carry irreversible or harmful consequences if ignored.

Examples. Thesis due Friday 11:59pm. No shellfish, allergy. $80 weekly food budget.

Conditional

True under a condition

Respected only when a specific condition is active, and ignored when it is not. The condition is the trigger.

Examples. No new writing after Thursday 6pm, active only in deadline weeks. Light Thursdays during crunch.

Soft

Real but bendable

Genuine preferences the AI respects by default, but can trade off when something harder conflicts. Hard wins, then conditional, then soft.

Examples. Prefers long uninterrupted writing blocks. Gets bored easily with repetitive meals.

06Interaction Grammar

The layout is the semantic model.

BLOOM is built on Shneiderman's direct manipulation principles. The context view is not a form to fill out. It is a spatial, manipulable representation of your world, where spatial properties are semantic, not just aesthetic.

Semantic zoom is navigation

The zoom level corresponds exactly to the data layer. World shows Parts, zoom into a Part to reveal Aspects, into an Aspect to reveal Points, into a Point to reveal Stories. No tabs, no menus. You zoom.

Drag-induced zoom

When linking across Parts, dragging a connector beyond a boundary zooms out to the World, and into another Part zooms in. The drag gesture itself controls the level of abstraction. Intent surfaces the right view.

Lateral linking, the graph layer

The hierarchy is a tree, but context is a graph. The same anxiety affects academics, gym focus, and budget. Aspects, Points, and Stories form explicit cross-Part edges that enable the AI to reason across domains.

Resize encodes weight, distance encodes relatedness

A larger node carries more context weight in the AI's prompt. Nodes placed closer together are more semantically related. Moving two Parts closer is a direct intervention on the AI's semantic model.

The World view: Parts laid out as bounded regions on a canvas, sized by weight.

07The Study

One claim, tested three ways.

The research tests a specific hypothesis: that giving users direct manipulation over the AI's context layer improves constraint survival under change, reduces context drift, and increases trust calibration. Three conditions isolate the contribution of manipulation over mere visibility.

C1

Baseline chat

A standard conversational interface. Context is buried in history. No visibility, no manipulation. We expect the lowest constraint survival.

✕Context visible

✕Context editable

C2

Read-only context panel

Context is externalized and visible alongside the chat. Users can see what the AI knows but cannot change it. Soft explainability without agency.

✓Context visible

✕Context editable

Hypothesis

C3

BLOOM, full

Context externalized, visible, and directly editable. Add, remove, edit, reposition, resize, and laterally link any node. Soft explainability with direct manipulation.

✓Context visible

✓Context editable

The core metric · constraint survival

Does a constraint set early in the conversation survive a pivot task midway through? C1 should lose the most. C3 should hold. C2 isolates what visibility alone buys you, before you add the power to act.

08The Product

BLOOM, built and running.

The architecture is implemented as a working React application. The chat lives on the left, the context view on the right, and semantic zoom carries you from World, through Parts and Aspects, down to the Stories in your own words.

World view

All Parts visible as bounded regions, sized by weight and arranged by relatedness.

Part view with aspect diamonds and points

Zoom into a Part

Aspects appear as diamonds. Points sit within, each carrying its constraint type.

Aspect view

Points with their grounding Stories and constraint chips, ready to inspect and edit.

The full BLOOM app with conversation and context side by side

Conversation plus context

Chat and the editable context model side by side, each updating the other.

09Contribution

A better contract between you and the AI.

Language models manipulate linguistic form without grounding in intent. BLOOM does not fix that. It gives the model a structured, externalized cheat sheet that you control, one that cannot be lost mid-context, overwritten by earlier outputs, or degraded through failed correction.

You remain the authoritative source of your own context. The AI reads it. The interface makes the relationship between the two legible and steerable. That is the contribution: not a better model, but a directly manipulable, hierarchically structured, spatially organized context layer that sits between you and the machine.

To be precise, what BLOOM is not

✕

Not a memory retrieval system

It does not fetch past conversations across sessions. That is a different problem. BLOOM keeps context accurate, visible, and steerable within the session as it evolves.

✕

Not a prompt engineering tool

Users do not write prompts. They manipulate context objects directly, and the AI reads those objects.

✕

Not a summary of the chat

The context view is a parallel structure, maintained alongside the conversation and updated by it, but structurally independent.

Giving AI a memory you can see, hold, and reshape.