Skip to main content
Memory GrainMemory Grain
GitHub
v1.5 · Standards Track · CC0 1.0

The Open Memory Specification
for Autonomous Systems

A family of three specifications that give AI agents immutable, portable, and verifiable memory — so they can remember, reason, and collaborate across systems.

Read the SpecificationView on GitHub
.mg

Open Memory Specification

The container format — immutable, content-addressed, cryptographically signed knowledge units.

Read OMS spec →RECALL

Context Assembly Language

The query language — non-destructive, deterministic context assembly from grain stores.

Read CAL spec →<fact>

Semantic Markup Language

The output format — flat, tag-based markup optimized for LLM context consumption.

Read SML spec →

How it works

Three specs, one memory.

OMS defines the container, CAL queries it, and SML renders it. Together they give any AI agent a complete, portable memory stack.

1STORE

Open Memory Specification

Defines the .mg container — content-addressed, immutable memory grains with SHA-256 hashing and COSE signing.

2QUERY

Context Assembly Language

A declarative query language for non-destructive, deterministic context assembly from grain stores.

3PRESENT

Semantic Markup Language

A flat, tag-based output format optimized for LLM context windows and human review.

In Practice

Any AI agent. One memory. Portable everywhere.

Imagine grains written by different AI systems — and read by others across industries, without any prior arrangement.

Slide 1 of 12: Fact grain by FitCoach
Written by
FitCoach
KineticAI

Learned from three months of workout data. Nutrition and scheduling agents use this grain to personalize recovery plans.

Fact · 0x01v1 · COSE Signed
{
  "type": "fact",
  "subject": "user:john-smith",
  "relation": "preferred_activity",
  "object": "morning_run_5k",
  "confidence": 0.94,
  "source_type": "pattern",
  "created_at": 1739980800000,
  "namespace": "fitness:preferences"
}
Read by
FoodBot
meal planning
CalendarBot
schedule optimization
FitCoach
training progression

Built for Enterprise

Trust at every layer.

Immutable by Design

Every grain is SHA-256 content-addressed. Once written, it cannot be altered — only superseded. Tamper evidence is built into the format.

Portable Everywhere

A .mg file is a self-contained container. Store it in S3, stream it through Kafka, carry it on a drive, or push it to Git. No vendor lock-in.

Audit-Ready

COSE Sign1 signatures, DID-scoped consent, and jurisdiction-aware retention. Built for GDPR, HIPAA, and SOX compliance from day one.

Conformance

Implement what you need

Declare your conformance level. Start minimal, add layers as requirements grow.

Level 1

Minimal Reader

Libraries, tools, verification scripts

  • Deserialize and verify grain blobs
  • Compute & verify SHA-256 content addresses
  • Field compaction (short keys ↔ full names)
  • All twelve grain types (0x01–0x0C)
  • Ignore unknown fields
  • Constant-time hash comparison
Level 2

Full Implementation

Agent frameworks, edge gateways

  • All Level 1 requirements
  • Serialize (canonical MessagePack)
  • Validate required fields per schema
  • Pass all test vectors
  • Multi-modal content references
  • Store protocol (get/put/delete/list)
  • Enforce invalidation_policy on supersession & contradiction
  • Atomic supersede operation (distinct from raw put)
  • Validate observer_type non-empty; emit oid/otype (v1.1)
Level 3

Production Store

Enterprise deployments, cloud platforms

  • All Level 2 requirements
  • Persistent backend (filesystem, S3, DB)
  • AES-256-GCM per-grain encryption
  • HKDF per-user key derivation
  • Blind-index tokens for encrypted search
  • SPO/POS/OSP index (hexastore) or equivalent
  • Full-text search (FTS5 or equivalent)
  • Hash-chained audit trail
  • Crash recovery and reconciliation
  • Policy engine with compliance presets
  • Partition Observation storage by observer domain (v1.1)

Blog

Deep dives with Raven & Trident

All articles →
memory-typesrecommendations

Grain Type Deep Dive: Recommendations

A comprehensive guide to the Recommendation grain type (0x0C), introduced in OMS v1.5 — governed, auditable proposals to change memory or agent configuration. Covers the propose/review/apply/roll-back lifecycle, why rec_status lives in the index layer, the normative dedup_key construction, applier execution authority, and reproducible undo.

17 min read
memory-typesskills

Grain Type Deep Dive: Skills

A comprehensive guide to the Skill grain type (0x0B), introduced in OMS v1.4 — the hybrid definition-plus-competence model, why the Fact + mg:capable_of convention was promoted to a first-class type, proficiency as a supersession chain, skill transfer between agents, and how Skills differ from Workflows and capability Facts.

15 min read
memory-typesconsensus

Grain Type Deep Dive: Consensus — Multi-Agent Agreement as First-Class Memory

How OMS Consensus grains record multi-agent agreement with threshold-based voting, trust-weighted confidence, and quorum protection — enabling auditable decisions across regulated industries, fleet coordination, and collaborative AI systems.

13 min read

Implement OMS today.

The specification is open and licensed under CC0 1.0. Read the spec, run the test vectors, and build interoperable agent memory in any language.

Read the SpecificationView on GitHub

OMS v1.5 · .mg Container Definition · Standards Track · CC0 1.0