The problem
Every fresh Codex session faces a bad choice: forget earlier work, or trust a transcript that may be stale, incomplete, or wrong. A useful agent memory must preserve continuity without quietly turning remembered text into truth or permission.
What I built
Integrity Seed is a local-first Codex plugin that gives each workspace evidence-linked continuity across sessions.
It starts empty by default. Installation imports no user history. It records deliberate handoffs and checkpoints instead of dumping full transcripts. Lifecycle hooks give a later Codex session a small, bounded memory capsule with exact source event IDs and an explicit coverage result. Recalled text stays untrusted. Current instructions and live evidence always win. Potentially write-capable tool use creates closure debt. A stale or replayed receipt cannot falsely mark newer work complete. Missing, corrupt, ambiguous, or duplicate hook delivery fails closed instead of being described as restored memory.
No API key, sample account, cloud service, or rebuild is required. The repository contains an installable deterministic release archive, setup instructions, a judge smoke demo, and the exact validation evidence.
The real demo
The video begins in a clean Linux account and empty Git project. Integrity correctly reports that it has no memory for DEMO-1. Codex then performs real work: it creates and verifies a file and records a handoff with the file hash and Action Log event ID.
A brand-new Codex session recalls that bounded event, cites the same source ID, and independently re-checks the live file. Then the file is changed outside Codex. A third session still remembers the historical event, but detects that the current file no longer matches it. The memory remains useful evidence without becoming false authority. That is the core GIGO defense.
How Codex and GPT-5.6 were used
GPT-5.6 Sol in Codex was used materially to inspect and refactor the original research system into a sanitized plugin, design adversarial lifecycle and replay tests, implement cross-platform hook wrappers, build a deterministic archive, run clean-room verification, and independently review the security boundary.
The human product decisions were equally important: local-first before networking, empty-by-default before importing history, evidence rather than authority, bounded capsules rather than transcript dumps, and explicit refusal to claim unproven fleet-scale safety or token savings.
The resulting 0.1.7-rc11 candidate passes 67 unit and contract tests from the exact release bytes. The verifier, local runtime, and handoff-to-recall demo passed on observed Windows 11 x64 and Linux x86_64 environments. A fresh Linux Codex task also observed exactly one SessionStart, UserPromptSubmit, and Stop lifecycle delivery with no duplicate project hook.
Why it matters
Integrity targets a problem that grows with every longer-running agent workflow: work becomes expensive to repeat, but dangerous to remember carelessly. Evidence-linked continuity can make Codex sessions easier to resume, audit, and challenge. This release proves the narrow local primitive; safe cross-host coordination and fleet orchestration remain future work, not demo claims.
Log in or sign up for Devpost to join the conversation.