Agent Memory System v0.2.0

Agent Memory System v0.2.0 makes the evidence-backed memory workflow easier to use and adds a bounded native Codex diagnostic.

Highlights:

• One-command Codex onboarding and a status command that report the current safe operator action.
• Review and promotion commands can resolve the current verified candidate generation without generated-path plumbing.
• A sealed, no-tools, 20-cluster Codex baseline/memory diagnostic with full local evidence replay and a privacy-safe aggregate receipt.
• Stronger verification of public-suite tasks, retrieval and injection provenance, execution order, artifacts, and raw Codex JSONL.
• New schemas and hosted Windows, macOS, Linux, race, fuzz, OpenCode, and public-history checks.

Observed on the frozen synthetic capability suite: baseline 1/20, memory 20/20, 19 wins, 1 tie, 0 losses. This is bounded evidence of benefit on that suite, not longitudinal continuous-learning certification.

Raw Codex events, messages, stderr, local paths, and the evidence ledger remain local.

Agent Memory System v0.1.0

Agent Memory System v0.1.0 is the first public release of a local-first, evidence-backed continuous memory system for Codex, Claude Code, and OpenCode.

Highlights:

• append-only local evidence capture with explicit loss and gap accounting
• episode reconstruction and evidence-backed candidate memory
• caller-attested review, promotion, supersession, and revocation
• readable, redacted portable memory with private Git synchronization
• scoped retrieval for Codex, Claude Code, and OpenCode
• encrypted evidence backup and fail-closed recovery
• reproducible Windows, macOS, and Linux archives for amd64 and arm64

The release is bound to protected main, seven required checks, and exact-archive acceptance on GitHub-hosted Windows and macOS runners. See PROVENANCE.json, REQUIRED_CHECKS.json, and the platform acceptance receipts attached to this release.

Known boundary: provider-hidden reasoning and data not exposed by an Agent runtime cannot be captured. Continuous-learning efficacy remains fail-closed until native Agent execution provenance and the fixed evaluation gates are satisfied.