Skip to content

CRBRO 2.10.0

Choose a tag to compare

@Octonove Octonove released this 18 Oct 10:31
· 73 commits to master since this release

CRBRO 2.10.0

Reliability, explicit evidence, project context and read-only diagnostics. Still 15 MCP tools, compatible with existing brains and callers.

  • Locked fresh keychain writes, ownership-aware locks and atomic context/session updates.
  • Durable shared-delivery outbox, idempotent retries and recovery archives when leaving a space.
  • Repetition no longer counts as verification; structured sources and explicit user/source verification.
  • Stable project identities, aliases, connection-local scope and structured pending items.
  • Separate relevance, provenance, freshness and verification in recall results.
  • Read-only doctor --json, frozen benchmark gates and expanded regression coverage.
  • Updated dependencies; release-time audit reported no advisories.

Validation: 943 tests passed, 6 skipped; Windows and Linux CI passed. Build and frozen retrieval/security benchmarks passed.

Installation

Use npx -y crbro-memory@latest in an MCP client, or download the .mcpb asset for Claude Desktop. Restart the client/server after updating; an already-running process does not upgrade itself.

Repository history notice

Two previously exposed credentials were removed from historical test fixtures. The WordPress application password was revoked and replaced; the old SSH password was confirmed rejected. All 46 previously published npm versions were checked for those two values and their tested encodings, with no matches. This is a targeted check, not proof against every possible secret or encoding.

The main branch and historical tags were rewritten. Existing contributors should preserve their own uncommitted work and use a fresh clone; do not merge the old history back. Copies in third-party forks or caches are not removed by rewriting this repository.

The operating-system credential vault is not included in releases. Local storage does not mean a secret can never leave the machine: explicitly retrieving it into an AI conversation or using it to authenticate sends it to the relevant client/service.