Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
148 changes: 148 additions & 0 deletions docs/architecture/intelligent-org-current-state.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,148 @@
---
title: 'The Intelligent Organization — Current State'
date: 2026-08-29
status: draft
tags: [architecture, intelligent-org, ai, memory, hypha]
---

# The Intelligent Organization — Current State

What intelligence is actually built on Hypha today, and how AI is used. Companion to
[What it is](../product/intelligent-org-features.md) (the target) and
[Design](./intelligent-org-design.md) (how to get there).

Snapshot of the `hypha-web` codebase as of 2026-08-29. Not a roadmap.

---

## One sentence

Hypha has a capable **in-app copilot** — onboarding, proposal drafting, space Q&A, navigation —
plus a **file/timeline catalogue** called Space Memory. It does **not** yet run the intelligent-org
loop: hear → remember → offer work → watch outcomes → revise beliefs.

---

## How AI is used

Stack: **Vercel AI SDK + OpenRouter** (default `openai/gpt-4o-mini`). Contract everywhere: **the
AI drafts, the member confirms.**

The system prompt’s north star is *THE AI DOES IT FOR ME* — propose, pre-fill, navigate; the
member reacts yes / tweak / no.

| Surface | What it does | Flag (default) |
| ------- | ------------ | -------------- |
| **Left AI panel** on space pages | Ongoing advisor: reads space data, drafts proposals, creates signals, opens the right screen | `enable-ai-chat` (on) |
| **Onboarding hero** | Conversational space creation — purpose, governance, visuals, nested spaces. Optional Live Voice (OpenAI Realtime) | `enable-onboarding-ai-hero` (on), `enable-onboarding-write-tools` (on), `enable-onboarding-voice-realtime` (off) |
| **MCP server** (`packages/mcp-server`) | Same tools for Cursor / external agents | always, stdio |

A model actually runs in those three places. Almost everything else that looks “smart” is
arithmetic, heuristics, or a template.

---

## What is shipped

### Space advisor chat

`packages/chat-server` — tools the model can call in a live space.

**Read:** space, members, documents/proposals, treasury holdings, signals, parent/child ecosystem
graph, org-memory catalogue, individual memory assets (text, PDF extract, images), web search.

**Write / act:** draft and pre-fill any on-chain governance proposal (one field at a time, then
open the form); create a Coherence signal; post to Human Chat as the member; summarize recent
Matrix discussion; ingest a call recording/transcript into Space Memory; navigate to the relevant
screen (`mcp_navigation`).

This is the main intelligence members feel: ask about the space, get a grounded answer, have a
proposal or signal drafted.

### Space Memory — L1 substrate

Closest thing to organizational memory that is **on this branch**. Coherence tab
(`enable-space-memory`, on). Aggregates:

- proposal uploads
- Matrix chat files/images
- call recordings and transcripts
- discussion summaries

The AI lists this via `get_org_memory_by_space_slug` and can fetch a file’s content. That is
**raw substrate** — chat, files, transcripts — not curated beliefs.

Discussion summaries are **not** model-written. They take the last few messages and truncate.

### Coherence signals — proactive, mostly not a model

A cron **signal orchestrator** watches memory ingest (new transcript, summary, upload) and may
emit a signal. Scoring is **arithmetic** — asset counts, title overlap, cooldowns, daily caps.
Copy is templated (*“Recent space-memory activity indicates a coordination opportunity.”*).

The AI *can* also create a signal from chat (`create_space_signal_by_slug`) and relay one to a
connected space. That path is model-driven.

### Onboarding AI

Conversational wizard: interviews for purpose and principles; infers voting / entry /
transparency; looks at other Hypha ecosystems and proposes a nested-space blueprint; generates
logo/banner images; creates the space on-chain after confirmation. Optionally continues in Live
Voice.

### MCP for external agents

Same catalogue as in-app chat: org memory, documents, people, treasury, signals, proposal
guidance, discussion summary, call ingest. Cursor or an IBA can use Hypha as context.

---

## Built, not on this branch

**Space Intelligence** (L3 beliefs) lives on `feat/org-memory` /
[PR #2461](https://github.com/hypha-dao/hypha-web/pull/2461):

- versioned Markdown artifacts in object storage (purpose, assessments, insights, …)
- human-approved publish (propose → member confirm)
- graph / sunburst of artifacts ↔ signals
- Hypha Energy starter pack
- IBA API-key write path

Until it merges, the AI has no always-loaded “what this org believes” corpus — only live queries
of Postgres / Matrix.

---

## Designed, not built

From [What it is](../product/intelligent-org-features.md) and
[Design](./intelligent-org-design.md):

| Layer / feature | Status |
| --------------- | ------ |
| **L1 substrate** | Partial: files, transcripts, and summaries ship as Space Memory. Chat message bodies stay on Matrix — no ingestion into Postgres (the design's long pole). |
| **L2 activity ledger** | `events` table exists and is barely used. Not the typed ledger. |
| **L3 beliefs** (org brief, objectives) | Built on `feat/org-memory`, not merged. No Shapers chat write path. |
| **L4 decision memory** | Does not exist |
| Projects / tickets / DRI / offer–accept | Do not exist |
| Hear → Think → Route agent | Does not exist. Chat is pull (member asks), not a listener on rooms. |
| My Work / All Work / Org homes | Do not exist |
| Reviews that write themselves | Do not exist |
| “Ask the org anything” with receipts | Partial: chat can query memory, no L3+L4 grounding or citation contract |

Scorecard from [Organizational Intelligence — Memory Architecture](./organizational-intelligence.md):

> L1 exists. L3 is built and awaiting merge. L2 exists as an unused table. **L4 does not exist.**

The signal orchestrator is the failure mode that document names: both “did something change?” and
“what does it mean?” done with arithmetic, so the output stays generic.

---

## Related

- [The Intelligent Organization — What it is](../product/intelligent-org-features.md) — the target
- [The Intelligent Organization — Design](./intelligent-org-design.md) — how to build it
- [Organizational Intelligence — Memory Architecture](./organizational-intelligence.md) — the four layers
- [Space Memory panel](../plans/space-memory-panel.md) — current L1 aggregation surface
- Space Intelligence spec — L3 as built on `feat/org-memory`; [PR #2461](https://github.com/hypha-dao/hypha-web/pull/2461)
Loading
Loading