Skip to content

Release Notes: v0.6.0

Date: 2026-10-14

Version 0.6.0 improves local resident request handling, project server setup, and session transcript capture. Release benchmark results are still pending; the transport measurements gathered during development are diagnostic and are not presented as release benchmark results.

Local resident and transports

  • Provider adapters use one resident owner per project and route requests through the shared router and MemoryService.
  • The client tries local IPC first, then the configured loopback HTTP listener, then the filesystem request queue when socket transports are unavailable.
  • Project installation creates .membrick/server.toml when it is absent. Its defaults configure the resident HTTP listener and dashboard URL. Existing settings are preserved. If the preferred port is occupied, installation selects and saves an available port in the configured range.
  • The filesystem queue uses directory notifications with recovery scans, bounded concurrent dispatch, and blocking workers for file operations. Request and response envelopes are transient; MemoryService continues to persist accepted writes before acknowledging them. A timeout after a request is claimed can have an uncertain outcome, so clients do not replay it under a new request ID.

See the local IPC architecture and HTTP and resident benchmarks.

Session transcript capture

  • Claude Code and Codex process complete transcript records incrementally and preserve observations in source order. A capture cursor advances only after the corresponding upsert is durably accepted, allowing retries after an interrupted capture.
  • Captured observations and session context use the existing SourceDocument and MemoryService write path. MemoryService journals accepted upserts before replying; background publication makes them searchable afterward.
  • Transcript shrinkage causes capture to reread and upsert from the beginning. Previously captured observations remain available as history.
  • Captured records retain their source and timestamp without depending on Git branch or revision checks.

See the session recording design, Claude Code hooks, and Codex integration.

Provider CLI and setup

  • Project workflows now use explicit subcommands: status, review, analyze, index, query, context, recall, inspect-index, session, and export. Running membrick shows help, and passing a folder path alone no longer starts a project scan. membrick status [PATH] reports existing index state without scanning project files.
  • Provider command families are consolidated around install, serve, verify, uninstall, hook, and statusline commands, with shared provider-specific configuration overrides.
  • Claude Code and Codex have uninstall commands that remove Membrick-managed integration settings while preserving project memory data and unrelated settings.
  • Older provider-specific flags remain accepted for compatibility. Routine help hides legacy tuning flags that are not intended as everyday options.

See agent setup and the provider integration guides.

Upgrade notes

  • Install or upgrade the 0.6.0 executable and run the relevant provider install command in each project to initialize missing project settings.
  • The default resident listener uses loopback. Keep .membrick/server.toml if you have configured a custom port or dashboard URL; installation preserves existing settings.
  • This release note does not claim a LongMemEval score or a release transport throughput figure. Those should be added only after the clean release benchmark runs are reviewed.