Menu
All work

Accessibility · Public interest

Bunch: a context system for continuity across memory gaps

I built Bunch to meet a continuity need that the tools I had tried did not address: recovering working context across memory gaps. The source code and data model are public to inspect, and a no-sign-in demo uses fictional records. The engineering problem was preserving context so the right information surfaces at the moment it matters. The accessibility problem turned out to be the same problem in a different hat.

My role
Product design, architecture, agent-assisted implementation, testing, and delivery.
System
Web app + MCP server, shared service rules, PostgreSQL.
Inspect it
Public source and a fictional-data demo. Source rights are reserved; this is not an open-source licence.
Diagram of Bunch’s data model linking people to hosting and fronting history.
Bunch’s published data model: people have pictures, hosting history, and fronting history; a return can have a catch-up with saved notes, tasks, decisions, and conversation summaries. Hosting records responsibility; fronting records presence.

Evidence

Proof you can inspect

Observed recovery
About 15 seconds for one real catch-up that had previously taken roughly 20 minutes to reconstruct manually.
Public implementation
Source code, setup documentation, and the data model are published.
Shared engine
The web app and MCP server operate on the same explicit records and service rules.
Production failure analysis
Nine OAuth failures connecting the MCP server to ChatGPT, each fixed with a regression test and without loosening the private-data boundary. Read it

The 15-second result is one personal observation, not a population-wide performance benchmark.

Engineering & product

What the artifact taught me

Under the hood

Bunch has two front doors onto the same records: a web application for browsing and editing, and an MCP server so an AI assistant can read and write under the same rules.

The record model is deliberately explicit. A missing entry means nothing was recorded; the system does not infer that nobody was present. That boundary is part of the product, not an implementation footnote.

The product insight

I started by thinking I was building identity-tracking software. The more general problem was continuity: preserve enough structured context that the next person, session, or interface can keep going.

That makes the architecture productizable across different surfaces. The audience-specific language can change while the core job stays the same: maintain canonical context, track changes over time, and hand a harness the smallest useful packet of truth.

The problem

I have spent years trying to accommodate a practical problem created by dissociative amnesia while maintaining a demanding professional career. Earlier tools helped me track who was fronting, but they were much better at looking inward than helping me recover what had happened in the outside world.

What I owned

I owned the product decisions, record model, web and MCP interfaces, and the rules that keep them consistent. I used AI as an implementation tool and retained responsibility for reviewing, testing, and shipping the result.

What I did

I experimented first with community tools like PluralKit and Simply Plural, then began building small applications against their APIs. That process taught me both API integration and the limits of identity tracking alone. Later, while working deeply with AI context management, I reframed the problem: the interface was not the important part. Continuity of context was. I began building Bunch as an accessibility harness designed around that idea.

What changed

During one weekend, another alter was continuously fronting and I had little usable memory of what had happened. Instead of spending roughly 20 minutes asking people and reconstructing events manually, Bunch gave me a summary of everything I had worked on since I was last present in about 15 seconds.

That changed the problem from “How do I keep track of who I am?” to “How do I make sure the next version of me can keep going?”

Have a problem in this neighborhood?

I’m happy to walk through the technical decisions, the parts that failed, or what I would build next.

Talk about the work See the other work
Privacy and site analytics