Context and Memory Workspace
A Cloudflare Worker workspace for conversation memory, search, and background AI jobs.
A workspace for long-running AI conversations, and the things I want to remember from them.
Jennifer Naomi Nguyen · systems project · Running privately · February 2026 – present · built with Claude Code and Codex
README · ARCHITECTURE · TECHNICAL · source
Why I built it
Long AI conversations lose continuity. The decision I made three weeks ago, the constraint that ruled out the obvious approach, the half-finished thread I meant to come back to — all of it is still in the transcript, and none of it is easy to pick up next time.
Context and Memory Workspace is the system I built for my own work. Conversations come in; summaries, wiki pages, a knowledge graph, and an evolving memory state come out.
What it does
- Streams conversations and stores the messages.
- Turns long discussions into summaries and Spell Book pages — a personal wiki I can return to.
- Searches conversations, wiki pages, branches, and themes by meaning.
- Builds connections between ideas as the conversation history grows.
- Supports conversation branches, files, images, and speech in and out.
- Runs the background work that keeps those pieces moving after I close the laptop.
There's also a small real-time experiment I call the water cooler. It started as a joke and became a useful way to see how shared state behaves when several participants are active at once.
How it fits together
The interface is React. A Cloudflare Worker connects it to D1, R2, KV, Queues, Vectorize, Workers AI, and a Durable Object. Each has a job: messages, files, memory state, background work, search, or shared live interaction.
The most useful split is between a conversation happening now and the work that can happen afterward. A streamed response takes one path; summarization, embedding, and wiki generation take another. The architecture page follows both.
Memory I can read
I wanted the results to be things I could open: a summary, a wiki page, a branch, a basin weight. That makes it easier to understand what the system carried forward and decide whether it's useful.
The technical page looks at where the memory lives, why summaries become search inputs, and how the knowledge graph grows. The source excerpts show the corresponding parts of the implementation.
Current status and next steps
I'm connecting the pieces around my own workflow and working through the tradeoffs as they appear. It brings together context, retrieval, background jobs, routing, and interface work in one system I use.
The workspace runs privately. These pages show its architecture, diagrams, and selected source. The next useful improvement is a visible status for background jobs, so a delayed summary is easier to investigate.
Connected projects
Attractor — Adaptive Context and Personalization focuses on memory and context delivery in a standalone prototype. Model Routing looks at how this workspace and the Multi-Provider LLM Orchestrator choose models and move work. Systems Logging and Telemetry looks at the signals those systems leave behind.