AI context and memory exploration

Context Capsules are a proposed UI/SDK layer that gives users direct control over what an AI remembers, when, and why. This exploration started as a personal whitepaper in August 2025 and evolved into a set of mockups for an MVP focused on developer use cases.

Why this exploration? #

LLMs can process millions of tokens but still forget what you asked an hour ago. They remember the name of your cat and forget the business-critical constraint you repeated three times. Memory systems shipped by major providers made things murkier rather than clearer — users don’t know what’s stored, what leaks between sessions, or how to correct it.

Prompt engineering is a band-aid. Context engineering is a step forward, but users still can’t see or control what’s being remembered. The cost is real: workflow friction, declining trust, and abandonment of otherwise powerful tools.

The proposal #

A Context Capsule is a lightweight, human-facing container for the stable signals an AI needs to behave well in a given domain:

Intent and use-case (“I use this AI for therapy, summarizing reports, daily writing”)
Stack and style (Python 3.11, Black, TypeScript + ESLint)
Constraints (“never rename variables”, “retain schema field order”)
Tone and verbosity (“no praise, minimal explanations”)
Mode — a behavior contract: Code Agent, Auditor, QA, Mentor

Capsules sit alongside automated memory rather than replacing it. The default behaviour stays seamless; user-selected capsules override, redirect, or reinforce model memory when it matters.

Design principles #

Effortless and frictionless. The UI layer must remain intuitive and minimal. Capsules integrate with existing interaction flows.
Intent as alignment. Capsules capture intent and inject it in relevant sessions to save the user time.
Identity through context. Capsules let the user seamlessly switch identities to tackle distinct tasks without leakage between them.

MVP focus: developers #

Developers waste hours repeating the same guardrails (“don’t rename variables”, “don’t touch other files”) across sessions and files. Capsules bind to a repo, auto-activate when the project opens, and expose a one-keystroke mode toggle (@auditor, @qa, @agent).

Two personas drove the MVP exploration:

Luke Jo — corporate maintainer by day, learner by night. Auditor mode on the financial pipeline at work, Mentor mode on his Space Invaders side project at night. Same person, two trust contracts.
Dr. Lee — clinical healthcare analyst. QA-only mode for cardiovascular risk review. No dosage autocomplete, no equation rewrites, no edits.

The mockups on the right show how a capsule surfaces these signals in a compact, scannable card.

Deliverables #

Whitepaper — full write-up of the problem, proposal, use cases, and MVP scope. Read the whitepaper.
Mockups — capsule cards for the two developer personas, included below.

My role #

Authored the whitepaper end-to-end: problem framing, proposal, design principles, use case mapping.
Designed the capsule data model and mode contracts (Allowed / Forbidden / Output schemas).
Produced the mockups for the developer MVP.

Status #

Exploratory. The whitepaper is published as open thinking under CC BY 4.0. The ideas continue to inform my work on AI tooling — including zapcoach.me and ongoing experiments with local and cloud LLM stacks.