Skip to content

OMPA

Universal AI Agent Memory Layer — Vault · Palace · Temporal Knowledge Graph

CI PyPI Python License

OMPA gives any AI agent persistent memory in one pip install. It combines three layers:

┌──────────────────────────────────────────────────────┐
│  Layer 1: Vault  — human-navigable markdown          │
├──────────────────────────────────────────────────────┤
│  Layer 2: Palace — agent-accessible metadata         │
├──────────────────────────────────────────────────────┤
│  Layer 3: Knowledge Graph — temporal SQLite triples  │
└──────────────────────────────────────────────────────┘

Verbatim storage plus local hybrid search (BM25 + embeddings). Measured on the full LongMemEval-S (n=470): R@5 0.979 for hybrid search and 0.970 for keyword-only search. See benchmarks/README.md.

Install

pip install ompa          # Core
pip install ompa[all]     # Core + local semantic search

Quick start

ao init
ao session-start
ao classify "We decided to use Postgres"
ao search "database decisions"
ao wrap-up

Why OMPA (vs raw Obsidian)?

For consulting leads: Why OMPA vs raw Obsidian? — one page on the agent-memory wedge, dual-vault hygiene, and the Locus / CognitionOS companion stack.

Companion stack

Layer Project Role
Memory OMPA Vault + palace + temporal KG
Retrieval Locus Vectorless, explainable RAG
Compliance CognitionOS Audit / RBAC / retention productization

What's next