LOAM — persistent, queryable memory for AI agents, kept as Linked Data, so knowledge accumulates across sessions.
Talk to an Architect Free Trial
Persistent Memory · RDF · Linked Data · Part of VOICE

LOAM

Memory that accumulates, instead of disappearing between sessions.

A queryable, RDF-based memory layer for AI agents: session history, entities, tasks, execution traces, and preferences, kept as Linked Data.

LOAM (Linked Open Agent Memory) is a queryable behavioral contract for AI agents, encoded in RDF-Turtle. It defines how an agent retrieves and applies standing instructions, with RDF memory files serving as the authoritative source — so prior interactions become reusable, machine-computable context that any agent or LLM can pick up.

Get it on GitHub How it fits VOICE
LOAM (Linked Open Agent Memory), layer by layerFive layers, top to bottom. Agents and LLMs: any agent and any LLM can read and write the memory. The retrieval protocol: load, select, apply, record. LOAM itself, a collective of six linked graphs: session, entity, task, trace, domain and preference. The Turtle files that hold it: core, preferences, ontology, index, how-tos and sessions. Virtuoso, which can load the files and query them over SPARQL, or the files can be read directly from the filesystem. VirtuosoLOAD IT, QUERY IT OVER SPARQLOR READ IT FROM THE FILESYSTEM Turtle filesCORE · PREFERENCES · ONTOLOGYINDEX · HOWTO · SESSIONS LOAMSESSION · ENTITY · TASKTRACE · DOMAIN · PREFERENCE Retrieval protocolLOAD · SELECT · APPLY · RECORDTHE AGENT FOLLOWS ITS OWN RULES Agents & LLMsANY AGENT · ANY LLMREAD AND WRITE MEMORY
What it keeps

An entity relationship graph collective

The memory is a collection of linked graphs rather than a transcript. Because each fact is RDF with a hyperlink identifier, it can be queried with SPARQL, joined to enterprise data, and reused by an agent that never saw the original conversation.

01 — Session Graph
Conversations, plans, goals
What was asked, what was planned, and what the user was trying to achieve.
02 — Entity Graph
People, organizations, products, assets
The things the work is about, each with a stable hyperlink identifier.
03 — Task Graph
Actions, tools, workflows
What was done, with which tools, in which order.
04 — Trace Graph
Execution traces, provenance
Where each result came from, for auditing, debugging, and reproducing it.
05 — Domain Graphs
Subject-specific knowledge
For example contracts and compliance: knowledge that belongs to one field of work.
06 — Preference Graph
User and agent preferences
Standing instructions an agent applies without being reminded.

Further domain-specific graphs can be added for any field of work.

How an agent uses it

Load, select, apply, record

01 — Load
Read the memory first
Before acting on a task, the agent lists the memory folders and reads core.ttl, preferences.ttl, the optional local-only private overlay, ontology.ttl, and index.ttl.
02 — Select
Pick the relevant context
The agent classifies the prompt’s intent and, when a SPARQL endpoint is configured, uses it to select the preferences, how-to documents, and recent sessions that apply.
03 — Apply
Follow the rule that applies
Where a rule points to a how-to document, the agent reads that document before acting, so the detail comes from memory rather than from recollection.
04 — Record
Write back what was learned
After a meaningful task, the agent writes valid RDF-Turtle to the memory: a session record, new entities, and any rule the user stated. Knowledge accumulates between sessions.
What lives where

A folder of Turtle files

The memory is plain files, so it can be read, versioned, and diffed like any other source. Each name below opens the file or folder in the repository. The files can also be loaded into Virtuoso and queried over SPARQL.

Memory file or folderHolds
core.ttlUser identity, agent identity template, and output-routing requirements
preferences.ttlThe public-safe hub: standing instructions as HowTo steps, grouped by topic
preferences.private.ttlAn optional local-only overlay whose assertions override the public defaults. It is never published; start from preferences.private.example.ttl
ontology.ttlVocabulary: prompt intents, retrieval policies, and context-routing rules
index.ttlA session index: one entry per session
howto/Companion specifications, one per sub-HowTo, linked from preferences.ttl
sessions/Episodic memory: one Turtle file per day, model, and agent environment
entities/The people, organizations, tools, and concepts the work mentions
scripts/Validation and utility tools, including the memory-protocol and session-record gates

Documentation and tools

FileHolds
README.mdFolder structure, the retrieval protocol, and the sub-HowTo map
AGENTS.mdThe retrieval sequence an agent runs before it acts, step by step
SESSION-START-HOOK.mdHow the memory is loaded automatically when a session starts
load_memory.pyThe loader: prefers SPARQL context selection and falls back to reading the files
In VOICE

The Recall and Enrich steps

VOICE pairs Virtuoso’s federated access with OPAL’s orchestration. LOAM is what lets the next agent session start from what the last one learned: it supplies the Recall step, preserving findings, relationships, and provenance, and the Enrich step, folding new results back into persistent context.

  • Portable. The context is not tied to one LLM’s conversation history, so models and agent frameworks can change without losing it.
  • Auditable. Execution traces and provenance are recorded as data, not buried in chat logs.
  • Governed. Public rules and a local-only private overlay are kept apart, so what is shared is a deliberate choice.
Get it

Start from the repository

RepositoryLOAM on GitHub ContextVOICE — the whole architecture OrchestrationOPAL — the harness

Knowledge that compounds

Give every agent session the context the last one earned.

Get it on GitHub Talk to an Architect