HN Showcase
Back to browse
Open Memory Specification (OMS), Context Assembly Language (Cal)

Open Memory Specification (OMS), Context Assembly Language (Cal)

by sathishmg·Mar 4, 2026·1 point·0 comments
Visit ProjectView on HN

AI Analysis

●●SolidBig BrainNiche GemZero to One

Content-addressed agent memory with append-only CAL—delete structural impossibility, not policy.

Strengths
  • •Delete-as-grammar impossibility forces genuine tamper evidence; elegantly solves audit/compliance via language design.
  • •Three-layer design (OMS container, CAL query, SML render) cleanly separates concerns; portable across Kafka, S3, Git.
  • •COSE signing + DID scoping + token budgeting shows thoughtful enterprise architecture, not hand-waved compliance.
Weaknesses
  • •No agent framework adoption yet—LangChain, CrewAI, AutoGen haven't integrated; spec is theoretical until frameworks ship support.
  • •10 grain types feel designed-by-committee; unclear if Belief/Consensus/Consent distinction survives real multi-agent systems.
Target Audience

AI/LLM framework authors, enterprise AI teams, regulatory compliance officers

Similar To

LLM memory layers in LangChain · Kafka event sourcing patterns · Git object model

Post Description

Hi HN. We wrote an open specification for portable AI agent memory.

The problem: every agent framework (LangChain, CrewAI, AutoGen) has its own memory format. There's no standard way to port memory between systems, verify it hasn't been tampered with, or prove to a regulator that you deleted it.

OMS defines three things:

1. *A binary container format (.mg)* — memory grains are content-addressed (SHA-256), immutable, deterministically serialized (MessagePack). Think .git objects for agent knowledge. 10 grain types: Belief, Event, State, Workflow, Action, Observation, Goal, Reasoning, Consensus, Consent.

2. *CAL (Context Assembly Language)* — a query language for assembling LLM context. The notable design choice: delete is a structural impossibility. Not a policy — the grammar has no production rules for destructive operations. Append-only writes, bounded execution, token-budget-aware assembly.

3. *SML (Semantic Markup Language)* — flat output format for LLM consumption. Tag names are grain types (`<belief>`, `<reasoning>`, `<consent>`). No XML processor needed.

The whole thing is CC0 (public domain). OWFa 1.0 licensed.

GitHub: https://github.com/openmemoryspec/oms

Happy to answer questions about the design decisions.

Similar Projects

AI/ML●●Solid

Assembly Language for Agents

This repo turns agent loops into tiny, explicit instructions (LF, LI, APP, INF, INT, BGT, OUT) so you can compose reasoning as discrete micro-prompts instead of one huge monolith. The example VIBE_CONTROLLER shows concrete ops for cognition, safety checks and branching — a neat way to formalize ReAct-style loops. It’s a clever, researchy DSL with a Rust interpreter, but it’s early-stage and sits in a crowded space of agent frameworks (needs runtime benchmarks, integrations and real-world demos to move beyond an interesting experiment).

Big BrainNiche Gem
vanilla-latte
224mo ago