AI-human Alignment and Cooperation: enabling AI to learn cooperation through long-term alignment

English | 简体中文

AICO, AI-human Alignment and Cooperation, is an open research framework exploring long-term human-AI collaboration.

It is not only concerned with preserving more historical information. It explores how AI can gradually learn to understand a person, understand relationships between people, and respond, collaborate, and act more appropriately across different goals and situations.

For AICO, long-term interaction is not a passive accumulation of chat history. It is an evolving alignment process: AI needs to know whom it is working with, whom it is currently interacting with, what the current topic and relationship mean, and how a multi-turn conversation should be organized and advanced. In PERSONAL mode, AICO serves the user and their relationship world. In EXPERT mode, it aligns with an expert's professional logic, service style, and decision process, while assisting interaction with clients.

---

Why AICO

Many AI systems can already preserve preferences, extract facts, retrieve history, or inject long-term memory into context. AICO aims to take a further step: make this long-term information participate in dialogue judgment and organization, so AI knows not only what happened before, but also how to collaborate with people appropriately now.

The Change AICO Seeks

Direction	The Capability AICO Pursues
01 From factual continuity to behavioral continuity	AI does not only remember that someone “prefers concise messages.” It also considers relationship stage, the current topic, and implicit interaction goals to choose a more appropriate tone, pace, and boundary.
02 From single-turn relevance to multi-turn strategy	When a conversation has stages and a purpose, AI does not simply answer sentence by sentence. It can organize the dialogue through a strategy tree: understand context, build common ground, enter the core request, and adjust its path from real feedback.
03 From static personalization to negotiated co-evolution	Alignment is not AI privately defining a person. In PERSONAL mode, it is calibrated through AI, the user, partner feedback, and interaction outcomes. In EXPERT mode, the expert has final authority over their logic tree, profile, and rules.
04 From fragmented memories to a relationship world model	Interactions with different people form an owner-private relationship graph. When talking with A, the system reads the local, relationally connected context around A rather than indiscriminately injecting all history.

From Alignment to Cooperation

flowchart LR
    I[Long-term interaction] --> A[Alignment model<br/>people, relationships, topics, goals, expert logic]
    A --> S[Strategy organization<br/>relationship subgraph, topic graph, strategy tree, RAG]
    S --> C[Cooperative interaction<br/>reply, recommendation, or professional decision]
    C --> F[Feedback and confirmation<br/>user, interaction partner, expert, system]
    F --> A

Loading

This loop turns long-term interaction into evidence for the next act of cooperation. AICO focuses on four questions: who is being aligned with, who is the current interaction partner, what the topic and multi-turn purpose are, and how AI can cooperate and evolve while respecting clear authority and confirmation sources.

---

Two Alignment Modes

Mode	Aligned Subject	Interaction Partner	Purpose
PERSONAL	The user themself	Self, friends, family, colleagues, other people	Learn the user's identity, preferences, memories, relationship network, and everyday dialogue strategies.
EXPERT	The expert	Client	Align with the expert's logic, style, knowledge structure, decision tree, and service workflow.

In EXPERT mode, the client is modeled as a service context for better expert-side replies. The client is not the long-term aligned subject.

---

What AICO Builds

Layer	Meaning	Used for
Personal Profile	Stable and dynamic profile of the user	Long-term personal alignment
Expert Profile	Expert style, school, knowledge structure, preferences	Expert logic alignment
Client Service Profile	Client case context, current need, risk signals, communication style	Supporting expert replies
Relationship Graph	Owner-private graph of people, relationships, events, constraints, and strategies	Reading relationship context before replying
Topic Graph	Dynamic semantic graph of extracted topics	Reuse, extend, or create strategy trees
Strategy Tree	Macro-level dialogue logic tree	Organizing multi-turn conversation flow
RAG Memory	Professional knowledge, personal memory, relationship memory, historical strategy	Compact retrieval for generation
Confirmation Records	AI/user/expert/system confirmation source	Traceability, revision, and governance

---

Owner-private Relationship Graph

AICO's relationship graph is not a global social graph. Each user owns a private graph built from their own conversations.

For user me, AICO may build:

me
├── A
│   ├── A's parent
│   └── A's colleague
├── B
│   └── B's mother
└── C

When me chats with A, AICO retrieves a local subgraph instead of the whole graph.

Retrieved Context	Why
Direct me-A edge	Relationship state, events, trust, tension
Person nodes for me and A	Profiles, preferences, constraints
Topic-relevant one-hop or two-hop relations	Background people and indirect context
Strategy implications	How relationship should affect the reply

AICO's rule of thumb:

The LLM proposes what should be updated; the system controls evidence, merging, deduplication, confidence, permissions, and persistence.

---

Current Features

Feature	Status
PERSONAL and EXPERT mode separation	Implemented in algorithm layer and frontend routes
Dynamic topic extraction	LLM-compatible extractor with local fallback
Topic graph reuse / merge / create	Implemented
Strategy tree runtime	First executable version
Owner-private relationship graph	First version implemented
Multi-source RAG	Supports knowledge, personal memory, relationship memory, strategy memory, expert profile, and client service profile
Expert/client frontend structure	April frontend retained and extended
Java alignment endpoints	Extended with AICO state, client service profile, and relationship edge details
Technical report	Coming soon

---

Repository Layout

aico/
├── api/                 # Shared protocols and AICOOrchestrator entry point
├── alignment/           # Topic extraction, topic graph, vector similarity, iteration jobs
├── perception/          # Profiles, personal state, and relationship graph
├── decision/            # Thought tree and strategy tree runtime
├── knowledge/           # Retrieval and multi-source RAG
├── evaluation/          # Response and feedback evaluation
├── generation/          # Prompt and response generation
├── backend/             # Spring Boot backend extended from April
├── frontend/            # Vue frontend extended from April
├── algorithm/           # Preserved and reconstructed algorithm assets
├── storage/             # JSON state store for current development
└── tests/               # Python tests

---

Quick Start

Python Environment

AICO uses a dedicated conda environment named aico.

conda env create -f environment.yml
conda activate aico
python -m pip install -e .

Run Python tests:

python -m unittest discover -s tests

Expected current result:

Ran 7 tests
OK

Python Example

from aico import AICOOrchestrator
from aico.api.schemas import AICOTurnInput, ClientMessage, InteractionMode

orchestrator = AICOOrchestrator()

output = orchestrator.process_client_message(
    AICOTurnInput(
        interaction_mode=InteractionMode.PERSONAL,
        counterpart_id="A",
        message=ClientMessage(
            client_id="user_1",
            conversation_id="conv_personal_A",
            text="I want to contact A again, but I do not want to pressure him.",
            metadata={"relationship_type": "friend"},
        ),
    )
)

print(output.response.text)
print(output.response.metadata["active_relationship_subgraph"])

Frontend

cd frontend
npm install
npm run dev

Route	Meaning
/personal	PERSONAL entry
/personal-workbench	Personal alignment workbench
/expert	EXPERT entry
/expert-chat	Expert workspace
/parent-chat	Lightweight client chat
/decision-tree	Strategy tree editor
/aico-alignment	Alignment state viewer

Backend

cd backend
mvn spring-boot:run

Method	Endpoint	Description
POST	/api/aico/alignment/turns	Record a turn and update alignment state
GET	/api/aico/alignment/users/{userId}/state	Get aligned subject state
GET	/api/aico/alignment/relationships	Get PERSONAL relationship state
POST	/api/aico/alignment/feedback	Record user or expert feedback

---

LLM Configuration

Topic and relationship extraction are designed around an OpenAI-compatible chat endpoint. If no endpoint is configured, AICO uses conservative local fallback logic so tests and offline development remain runnable.

$env:AICO_LLM_ENDPOINT="http://localhost:11434/v1/chat/completions"
$env:AICO_LLM_API_KEY="local-dev-key"
$env:AICO_LLM_MODEL="your-local-model"

Embedding currently uses a deterministic local implementation for development and tests. It can later be replaced with a production embedding provider.

---

Roadmap

Area	Next Step
Technical report	Release AICO technical report and architecture diagrams
Relationship graph UI	Click an edge to inspect people, events, evidence, constraints, and confirmation status
Strategy tree executor	Tighten node transition control and bind traces to real conversations
Expert confirmation workflow	Complete candidate topic/tree/node confirmation in expert workbench
Persistent backend	Move JSON state toward database or event-stream persistence
LLM extraction	Replace local fallback with stronger structured extraction and production embeddings

---

License

AICO is released under the GNU Affero General Public License v3.0 (AGPL-3.0).

---

Citation

If AICO contributes to your work, please cite the project using the entry below. Until the technical report is published, this software citation is the recommended reference.

@software{aico_lab_aico_2026,
  author = {Pengcheng Zhou},
  title  = {AI-human Alignment and Cooperation: enabling AI to learn cooperation through long-term alignment},
  year   = {2026},
  url    = {https://github.com/PKQZPC/AICO},
  note   = {Open-source research framework}
}