Ennoia

User model layer across AI systems.

Internal codename: This project is developed under the codename "cortex". All source code internals, project management documents, and the local directory (~/cortex) retain the "cortex" name. The external brand and CLI command are "ennoia".

Status

Stage 25 — Cloud Relay Architecture Research Complete

Goal

Build a persistent user model that influences AI behavior across tools.

Ennoia is not a memory DB or RAG layer. It aims to maintain a structured, versioned representation of the user (projects, goals, preferences, constraints, skills, states, decision rules) and feed it back into AI systems to influence prompt construction, planning, and tool selection.

Layout

src/
  index.ts              minimal entry point
  core/
    user-model/         user model schema and storage
    memory/             raw memory events
    runtime/            runtime context assembly
  ingestion/            transcript / event ingestion
  extraction/           structured candidate extraction
  adapters/             per-agent adapters (e.g. Claude Code)
scripts/                one-off operational scripts

Requirements

• Node.js 20+
• npm 10+

Usage

npm install
npm run typecheck  # type-check without emitting
npm run build      # compile to dist/

30-second Demo

Install ennoia, then run in any Claude Code project:

ennoia sync --from claude-code

You'll see candidates extracted from your workspace — README, CLAUDE.md, .claude/agents/, package.json. Confirm the ones that are actually about you.

Then:

ennoia inject --format text

Copy the output. Paste it into ChatGPT or Claude web. Ask them anything.

You'll notice: they'll act like they know you — your projects, your preferences, your constraints. Not because they remember you, but because Ennoia gave them a model of who you are.

This is not memory. This is you, becoming a first-class entity across AI systems.

CLI

The ennoia CLI is exposed via bin/cortex. You can invoke it directly or run npm link in this directory to put ennoia on your PATH.

ennoia save "I'm working on the Ennoia project"
ennoia save "avoid single point of failure"
ennoia context

Expected output for ennoia context:

[User Context]

Projects:
  (none)

Goals:
  - I'm working on the Ennoia project
  - avoid single point of failure

Preferences:
  (none)

Constraints:
  (none)

Decision rules:
  (none)

User model data is stored at ~/.cortex/user_model.json. The file is created on first use. Edit it directly if you want to adjust fields outside of save.

Inject (text vs structured)

ennoia inject produces an agent-facing payload from the user model. It supports two output formats:

ennoia inject                                     # default: text
ennoia inject --agent claude-code                 # text, agent-tuned
ennoia inject --format text                       # explicit text
ennoia inject --format json                       # structured pack (JSON)
ennoia inject --agent claude-code --format json   # both flags

• --format text (default) returns the rendered instruction string. Stable for prompt concatenation and human inspection.
• --format json returns a Structured Injection Pack v0.1 — a stable JSON object that adapters / scripts can consume directly without re-parsing the rendered text.

This is not a runtime auto-injection mechanism. Ennoia does not attach itself to any agent process; it only produces a payload you (or an adapter) decide what to do with.

Structured Injection Pack v0.1

Top-level shape:

version              "0.1"
generated_at         ISO-8601 timestamp
source               { generator, generator_version, user_model_schema_version,
                       agent, selection_strategy }
user_summary         { total_entries, counts: { <kind>: number } }
entries              authoritative list of confirmed user-model items
projects | goals | preferences | constraints | skills | states | decision_rules
                     derived per-kind views over `entries`
open_questions       reserved for future scoped/task-aware selection (always [])
instructions         { text, notes[] } — restrained guidance for downstream agents

Each entries item carries:

id, kind, content, provenance, confirmed, created_at, updated_at, details?

details only contains fields meaningful to the entry's kind (e.g. when/then for decision_rule, level for skill). Empty user-model categories produce empty arrays — never missing fields or null.

The pack reflects the current confirmed user model only. It is not a ranking, not a scoped selection, and does not claim relevance to the caller's current task.

Project records

Formal project management documents (spec, decision log, stage archives, reviews) live in the companion repository:

~/ai-project-os/2_projects/cortex/

CLAUDE.md in this repo is local runtime context only and is not checked in.