Skip to content
alexk202Public

About

Common path from policies to implementation for long-term AI memory

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

MAQS-Way — Common Path from Policies to Implementation

DOI

Status: v0.3 — canonization release (all 12 known gaps from v0.2 closed; open known gaps: none) Core: MAQS v1.9 remains audit core: https://github.com/alexk202/memory-architecture-standard

Formula: MAQS = practical audit of architecture and typical failures (how to verify) Way = common model + measurable criteria + lifecycle + reproducible verification (how to build)

MAQS checks that mechanism exists. Way proves it works.


Why Way?

  • Policy layer is overcrowded: GDPR Art17, EU AI Act Art12/13/19, RBAC, governance docs
  • Audit layer exists: MAQS v1.9 — rare checklist backed by 20k+ tracebacks and production incidents
  • Path layer was missing: which tools, patterns, schemas to implement policy in memory architecture and how to prove it

Without path, policies are written by lawyers, audits are done ad-hoc by engineers.

What Way adopts (not competes)

Family What it provides How Way uses it
OMPS rdh073/omps Vendor-neutral portability memory-item.schema.json Canonical structural contract
AMSP sanej/amsp Lifecycle + scope model active→deleted, user-portable etc Scope + state machine
OAMP deep-thinking-lab Provenance mandatory agent_id binding §4.3 Provenance as requirement
Zep/Graphiti getzep/graphiti Bitemporal valid_at/invalid_at + t_created/t_expired Lifecycle + invalidation
Tombstone poojan6216/tombstone Physical deletion verification 92-100% residue bench Deletion lattice
LoCoMo/LongMemEval/MemoryAgentBench Measurable quality R@k, temporal, abstention Golden sets
OWASP Agentic ASI06 Memory poisoning tests Security verification
EU AI Act Art12 + OTel GenAI Logging by design Observability

Three layers

1. Principles — 8 immutable: Veracity, Provenance, Controlled Aging, Isolation, Deletability, Recoverability, Portability, Measurability

2. Architecture — Data model (OMPS + OAMP), lifecycle ACTIVE→STALE→CONTESTED→SUPERSEDED→EXPIRED→ARCHIVED→DELETED→TOMBSTONE→COMPACTION, bitemporal with temporal priority (invalid_at < now → EXPIRED, not SUPERSEDED), 6-tier deletion cascade graph_edges → nodes_if_sole_provenance → embeddings_fk → facts_fk → artifacts → rounds, Content-as-Data fenced <retrieved_context>, visible_to semantic masking

3. Verification — Property / Verification / Example separated. Golden sets + metrics CI gate / target: R@5 ≥0.7 / ≥0.85, temporal ≥0.65 / ≥0.8, abstention ≥0.8 / ≥0.9, duplicate ≤5% / ≤2%, contamination ≤1% / 0, residue 0. Deletion lattice LOGICAL/PHYSICAL/SEMANTIC/MODEL. EU AI Act logging + OTel with allowed/forbidden keys. Canon: maqs-way-standard.json → verification_suite.metrics.

Repo structure

SPEC.md — main spec v0.3 EN
SPEC-RU.md — RU version v0.3
maqs-way-standard.json — machine-readable v0.3.0 (8 principles P1-P8)
conformance/ — valid/invalid vectors + manifest.json
evals/ — how to run LoCoMo/LongMemEval
docs/ — mapping to MAQS, landscape, outreach drafts, articles
proposals/ — v0.1 roadmap snapshot (historical)

Companion artifacts

Self-audit evaluator, cheatsheets (P1–P8 + worked patterns), wired RED self-tests, GitHub Action template: https://github.com/alexk202/maqs-way-evidence — derived from canon, nothing there rewrites it.

Relation to MAQS

  • MAQS v1.9 = audit checklist you put on table, CRIT/IMP/REC verdict, backed by incidents
  • Way = governance contract that references MAQS sections 0-J as implementation evidence

Two repos, one coordinate system. MAQS stays clean.

Quick start

# inspect the machine-readable standard (valid JSON)
python -m json.tool maqs-way-standard.json
# conformance corpus (expected verdicts + rule_ids)
cat conformance/manifest.json

Audit CLI v0.1 ships (schemas/maqs_way_audit.py: schema + manifest + lint + coverage, namespaces, canon-pin) — run python3 schemas/maqs_way_audit.py all. data_model.$ref is active (schemas/memory-item.schema.json landed, gate green). Executable behavioral harness — planned, v0.4 (see SPEC §6 Backlog). schema_version (omps-0.1.0+way-v0.1.3) pins the data model, not the release (v0.3): the model is frozen since v0.1.3, the release grows — by design.

See docs/articles/devto-draft.md for the first DevTo article (v0.1.4). Both articles carry a dated update (2026-10-29) pointing to v0.3. v0.2 article (live): https://dev.to/aleksandr_kossarev_e23623/maqs-way-v02-shipping-a-standard-with-open-gaps-and-naming-them-5cef Worked audit examples: Mem0 · Graphiti · Letta.

Check not that the mechanism exists. Check that it works.

About

Common path from policies to implementation for long-term AI memory

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages