Learned by building, not by studying from a distance.

This isn't a research programme run apart from the work — it's action learning: the tooling problems below only get scoped, tested, and revised because a real build, the ARC Community Agent (see the hackathon), keeps forcing them into the open.

None of these six threads is the destination. Each is infrastructure for agents — and the organisations behind them — to collaborate without losing context, provenance, or trust. Open questions are tagged in the source and welcome challenge.

[ 01 ]

Tool Institutional intelligence

Transcript-native institutional intelligence.

Conversations as the source of truth an agent needs before it can represent an organisation to anyone else.

Building institutional memory from transcripts is not primarily a tool-calling problem; it is a memory-compilation problem. Conversations are high-volume, low-signal raw material that needs to be turned into a temporal graph of episodes, claims, contradictions, and unresolved questions — something a future agent (or human) can actually reason over, not just retrieve from, before it speaks for an organisation to the rest of a network.

→ research/The Coherence Company/agent-orchestration-and-memory-systems.md

[ 02 ]

Tool Harness design

Agent runtimes & the harness split.

The 2026 convergence: harness, execution, memory — three planes a collaborating agent needs, not one chatbot.

Serious agent systems are separating the control plane (turns, approvals, tracing) from the execution plane (sandbox, files, ports) from the memory plane. We compare live runtimes — OpenClaw, HERMES, LangGraph, the OpenAI Agents SDK, and Microsoft Agent Framework — against transcript-first institutional workloads, including the ARC Community Agent's.

→ research/The Coherence Company/coco-openclaw-vs-hermes.md

[ 03 ]

Tool Consolidation

Dreaming & consolidation architectures.

Periodic sweeps that promote evidence into durable knowledge — wake, sleep, replay.

The frontier of agent memory is no longer "store more"; it is write-manage-read with explicit consolidation, update, and forgetting policies. We track the lineage from Dyna and DreamCoder through Generative Agents, Reflexion, ReasoningBank, ReMe, and OpenClaw's core Dreaming subsystem — looking for trustworthy, governed consolidation an agent can lean on when it is synthesising across organisations, not poetic generation.

→ research/The Coherence Company/openclaw-dreams-and-hermes-dreaming-architecture.md

[ 04 ]

Tool Knowledge graphs

Compiled memory & temporal graphs.

Pages, not chunks. Provenance, not retrieval — because provenance is what lets an agent attribute a claim to the right organisation.

The cleanest memory model is not "retrieve better chunks" but compile raw sources into a persistent, interlinked wiki that compounds over time. We benchmark SurrealDB, Neo4j Agent Memory, Zep / Graphiti, Letta, and Mem0 against the requirements for compiled, evidence-backed institutional knowledge with temporal facts and provenance — the shape a source-linked, cross-organisational store needs to take.

→ research/regentribe/surrealdb-alternatives-research.md

[ 05 ]

Tool Commons infrastructure

Knowledge infrastructure for distributed communities.

GitHub for authoring, queryable graph for retrieval, MCP for any tool — the closest existing pattern to what a cross-organisational agent needs.

Genesis Brain Light is a two-layer system: members edit markdown in their own AI tools and push to GitHub; an automatic pipeline lifts those changes into a SurrealDB knowledge graph exposed via MCP. The aim is a knowledge commons that any AI assistant can query, regardless of which tool the human — or the organisation — is working in.

→ research/regentribe/genesis-brain-light-design.md

[ 06 ]

Tool Publishing medium

The Living Book.

Knowledge that grows, not knowledge that ends.

A new medium for deep knowledge — where the author maintains authoritative coherence, the community adds annotation and extension, multiple coherence streams coexist as different lenses, and contributors are recognized for the value they add. It is the publishing model behind ARC's build-in-public feed: a record several organisations can annotate without losing whose voice is whose.

→ research/The Coherence Company/living-book-manifesto.md

← Back to Lab №1