Attractor — Context Research - Architecture
The research uses the shared Attractor tool, with a comparison path alongside the ordinary memory-update path.
README · ARCHITECTURE · TECHNICAL · source
Start with the same conversation and state
The compare command loads a transcript and the current state. Each comparison leg chooses an engine and model, summarizes the conversation, and generates a proposed update. The resulting summaries, deltas, connections, and patterns are shown together.
Comparison results are recorded in the local run log without applying them to the live memory state. Ordinary ingestion applies the update and records its output through the ingestion path.
Each leg follows transcript + starting state → model summary → proposed update → comparison output and run log.
Shared pieces
| Piece | Role in the research |
|---|---|
| Engines | Call a model through the HTTP API or local Claude CLI |
| Prompt builder and parser | Build the update request and interpret the returned structure |
| Update core | Apply the same bounds and decay rules to proposed changes |
| Store | Supply the state the comparison begins from |
| Local run log | Group successful outputs by conversation hash and model |
The shared architecture and state rules document those pieces in one place.
Extending the research
The planned study adds a second view of the conversations: concept extraction, embeddings, and clusters tracked over time. Comparing that view with Attractor's state would help investigate whether the same patterns appear outside its own labels and update rules.
A later injection comparison would also record whether a context block was delivered at session start and exactly which block it was. These additions are described in the study plan; they are not part of the completed comparison examples.
Snapshot distinction
The current snapshot includes comparison outputs for specific conversations and model choices. Independent pattern analysis and memory-injection comparisons remain planned work.