Instead of pasting context into every AI chat, one project listens in the background

Every new AI conversation starts with the same tax. The user pastes a Slack thread, links a pull request, and re-explains the decision a meeting already settled, so the model can rebuild an understanding that already existed yesterday. Multiply that across a team that ships daily, and the cost shows up in minutes lost and in prompts that read like institutional memory.

Twin, a single-author open-source project, is testing whether that cost can be moved out of the prompt. It sits between a team's existing tools and any language model, continuously ingesting activity from Slack, GitHub, meetings, and docs, and turning it into what the project's creator calls a "situation model": a persistent, model-independent layer of computational understanding that the next conversation can read from instead of reconstructing.

Twin's mechanism rests on a single claim: cognitive state should live outside the model. Twin does not retrieve documents, index files, or stand in for an agent. It observes the team's stream, correlates across sources, and runs a "reflection" step that revises the model as new activity arrives. The author positions the project explicitly outside the current categories: it is not retrieval-augmented generation, not a knowledge graph, not an agent framework, and not a chatbot. The repository exposes the model through an MCP server (Model Context Protocol, an open standard for connecting language models to external context), which automatically injects the relevant slice into each new request. The cognitive state lives outside the model and survives across sessions.

The first public demo ran on Claude Sonnet 4.6. A fresh Claude conversation was opened with no custom memory, no project rules, no prompt describing the repository, and no access to local project files. Only Twin's MCP server was in the loop. Asked to explain why a specific feature had become a launch blocker, which pull request had resolved it, and what the project's new state was, the model answered correctly without ever receiving the underlying Slack threads or PR raw data. A YouTube walkthrough of the same test is also available, and the package itself ships on PyPI as twin-cognition, confirming the packaging and demo surfaces exist independent of the Reddit announcement.

One developer, one public software project, one model, and one creator-reported test make up the entire evidence base. There is no published benchmark, no peer review, and no claim that the approach generalizes to other teams, codebases, or model families. Treating the result as evidence that "AI now has continuous memory" would over-read a single first-milestone test the way treating one face-recognition demo as proof of general vision would. The honest read is narrower: under one specific configuration, a cognitive layer can answer project-state questions that a fresh conversation with no local context could not.

The current race in AI-assisted work is over which model carries the most context: longer context windows, better retrieval-augmented generation, project-memory features bolted into chat products. Twin's bet is that the model does not need to carry the context at all, because a separate, model-independent layer can serve it. If that separation holds, the design question stops being "how do we fit more tokens in" and becomes "how do we keep a persistent model of the work itself." That is a different layer than retrieval and a different dependency than the model. The second-order effect is on the model market itself: a cognitive layer that any model can read from commoditizes the model's own memory features and shifts competition back to raw capability, latency, and cost.

The falsifier for that bet is straightforward. A cognitive layer that observes, correlates, and reflects is more software to build, more state to keep coherent, and more failure modes than a longer context window. A model vendor that simply expands its context to cover the same surface could win on simplicity. The next milestones to watch are whether the approach holds on a second, independent project, whether the reflection step is reproducible across runs, and whether the repository picks up outside contributors.