memgov — cross-agent memory governance

Gate what an AI agent's persistent memory may assert before the agent relies on it. Local-first, deterministic, stdlib-only (Python 3.11+).

Memory is a point-in-time snapshot, not live state.

memgov transplants the RCGov axiom — InjectContext_t ⇒ ContextReady_t, govern what a model reads before it answers — onto the memory layer of coding agents (Claude Code, GPT Codex, and anything else whose memory is Markdown on disk).

The failure it prevents

Two agents share a workspace. Agent A's memory says a paper is "not yet deposited". In reality it was published yesterday — by a session that didn't update A's memory. A, trusting its memory, deposits it again.

Multiply that by every status-bearing claim agents write down: "awaiting review", "CI is broken", "tool X can't start". Each goes stale silently the moment another session, another agent, or the human changes the world.

What memgov does

1. SHARED_STATE.md — a small, human-readable table of committed canonical live state: fact, current state (dated), verification anchor, and a key regex. The session that changes durable state overwrites the matching row — the row overwrite is the cross-agent handoff.
2. memgov scan — cross-store drift detection (the CONFLICT_MAP): flags memory lines that match a resolved row's key while still carrying a stale/status marker. Non-compensatory: flagged lines are a review queue, never auto-deleted; only verification clears them. Every pass appends to a measurement ledger (LEDGER.jsonl) — clean passes are evidence too.
3. memgov fresh — freshness check: status-bearing memory files untouched for N days get flagged before an agent relies on them.
4. memgov hook-context — one command that emits the SessionStart context block (SHARED_STATE + freshness + drift summary) in the hookSpecificOutput JSON schema accepted by both Claude Code and Codex hooks unchanged.

Quickstart

pip install .            # or: run in-place with python3 -m memgov
cd /path/to/your/workspace
python3 -m memgov init   # scaffolds memgov.toml + SHARED_STATE.md
$EDITOR memgov.toml      # point [stores] at your agents' memory dirs
python3 -m memgov scan --summary
python3 -m memgov hook-context   # what your agents will see at SessionStart

Then wire the SessionStart hooks:

• Claude Code: merge adapters/claude-code/settings.snippet.json into ~/.claude/settings.json; optionally install adapters/claude-code/skills/memory-restore/ as a skill.
• Codex: merge adapters/codex/hooks.snippet.json into <repo>/.codex/hooks.json; paste adapters/codex/AGENTS.section.md into your AGENTS.md.

Discipline (the part software can't do for you)

• Status-bearing memories get an absolute date and a verify anchor (a command / path / API that yields ground truth).
• The session that changes state updates SHARED_STATE in the same session. The tooling is the safety net, not a license to write stale notes.
• Agents never edit each other's private stores; corrections flow through SHARED_STATE rows. See docs/MEMORY_GOVERNANCE_BASIS.template.md for the fill-in governance document.

Design lineage

memgov is the memory-domain sibling of RCGov (context governor) and a cousin of mmv-secretary (session continuity clerk). Concept mapping: SHARED_STATE rows ≈ RCGov's Authority Commitment Gate; scan output ≈ CONFLICT_MAP + Non-Injection Report; the review-queue rule ≈ non-compensatory gating.

What it deliberately is NOT

No embeddings, no LLM calls, no background daemon, no writes to your memory stores (except the opt-in .last_restored marker). Detection is regex over Markdown — conservative, auditable, tunable in memgov.toml.

Development

python3 -m pytest tests/ -q   # 10 tests, stdlib + pytest only

Citation

A companion systems paper documents the motivating incident, the RCGov theory transplant, and the first measured detection-and-repair cycle:

Toeda, T. (2026). Memory Is Not Live State — Cross-Agent Memory Governance by Transplanting the Reflective Context Governor onto Agent Memory (memgov v0.1.0). MOBIUS LLC. DOI: 10.5281/zenodo.21231158.

See CITATION.cff for machine-readable metadata.

License

Code: AGPL-3.0-or-later. Documentation under docs/: CC BY-NC-SA 4.0 (see docs/LICENSE-DOCS.md).