gwt-context is a Global Workspace Theory-inspired MCP server: facts compete for a bounded workspace, broadcasts fan out to independent processors, and accepted proposals drive the next memory actions.
The current runtime has a bounded workspace, ignition threshold, recurrent link activation, bus arbitration, subscriber budgets, and MCP tools for exact memory operations.
Candidates enter a preconscious buffer, compete through specialists, and only high-activation winners become globally visible in the workspace.
Structured resolve, semantic recall, relation continuation, contradiction checks, and follow-up critique read the same broadcast and propose actions.
Backup, restore, reset, dedup import, working-memory compaction, exact collection queries, and admission reasons are available through MCP.
Latest bounded sanity run was executed on 2026-04-29 against the configured OpenAI-compatible Qwen endpoint, with bus-on and bus-off slices. Numbers below are the artifacts of the run that ships in this release candidate.
Use hash embeddings for deterministic local checks and MCP client smoke tests without downloading embedding models. Bounded sanity, RULER, and LongBench all pass under the hash provider.
Start with the honest GWT report before repeating the GWT claim. It names what is real in the runtime and what still needs independent processor loops.
Claim and limits: what is real in the runtime, what still needs independent processor loops.
MCP and Qwen evidence: the run that produced the numbers above.
Gates and commands required before tagging a release candidate.
Current Qwen sanity matrix, bus deltas, release gates, and the latest delta example.
State boundaries: workspace, bus, subscribers, and memory operations.
Payload shapes for every tool exposed through MCP.
Current release candidate: what changed, what is pending.
Tag-specific release candidate notes and final tag checklist.