An Agent Zero community plugin that integrates Cognee, an open-source AI memory platform combining vector search with graph databases, into Agent Zero's memory lifecycle.

Plugin Name: a0_cognee (underscore) | GitHub: https://github.com/neurocis/a0_cognee | Version: 1.0.0 | License: MIT

Cognee transforms raw conversation data into a connected knowledge graph. This plugin automatically:

• Retains — Extracts key knowledge from conversations and stores it in Cognee
• Recalls — Enriches agent responses with semantic and graph-based search results
• Contextualizes — Injects relevant knowledge graph context into the system prompt

Unlike simple vector-only memory, Cognee links documents, entities, and concepts into a graph structure, enabling relationship-aware retrieval with 14+ search strategies.

Agent Zero                          Cognee Server (Docker)
├─ monologue_start                  ┌──────────────────────┐
│  └─ _20_cognee_init               │  REST API :8000      │
├─ system_prompt                    │  ├─ /api/add         │
│  └─ _30_cognee_context  ────────► │  ├─ /api/cognify     │
├─ message_loop_prompts_after       │  ├─ /api/search      │
│  └─ _51_cognee_recall   ────────► │  └─ /api/v1/datasets │
├─ monologue_end                    │                      │
│  └─ _52_cognee_retain   ────────► │  Triple Store:       │
└─                                  │  ├─ Relational (meta)│
                                    │  ├─ Vector (semantic)│
                                    │  └─ Graph (relations)│
                                    └──────────────────────┘

• Agent Zero v0.9.x or later
• Cognee Server running and accessible (Docker recommended)

# Quick start with Docker
echo 'LLM_API_KEY="your_openai_key"' > .env
docker run --env-file ./.env -p 8000:8000 --rm -it cognee/cognee:main

# Or with Docker Compose
git clone https://github.com/topoteretes/cognee.git
cd cognee && cp .env.template .env
# Edit .env with your LLM_API_KEY and other settings
docker-compose up -d

Verify it's running:

curl http://localhost:8000/api/health
# {"status": "healthy"}

1. Clone the plugin into your Agent Zero plugins directory:

   cd /path/to/agent-zero/usr/plugins/
   git clone https://github.com/neurocis/a0_cognee.git cognee

   Note: The plugin is named a0_cognee in the community index (underscore format for CI compliance), but Agent Zero installs it locally as cognee (without the prefix).

2. Enable the plugin in Settings → Plugins.

3. Configure your Cognee server:

   • Required: Set Cognee Base URL in Settings → Cognee Knowledge Memory (e.g., http://your-cognee-host:8000)
   • Optional: Set COGNEE_API_KEY in Settings → Secrets (for authenticated instances)

Once available in the Plugin Hub, install directly from Settings → Plugins.

All settings are configurable via the plugin's WebUI panel (Settings → Plugins → Cognee Knowledge Memory).

At the end of each agent monologue:

1. The conversation history is sent to a utility LLM with an extraction prompt
2. Key facts, relationships, and knowledge are extracted as structured data
3. The extracted knowledge is sent to Cognee via POST /api/add
4. If Auto-Cognify is enabled, POST /api/cognify builds the knowledge graph
5. All retention runs in a background thread to avoid blocking responses

During each message loop iteration:

1. The user's message is used as a search query
2. Cognee is queried with the configured search type
3. Results are formatted and injected into loop_data.extras_persistent
4. The agent sees these results as additional context when generating responses

When the system prompt is assembled:

1. Recent conversation history is used as a contextual query
2. Cognee returns relevant knowledge graph connections
3. Results are injected into the system prompt via the cognee.context.md template

Datasets are automatically scoped per project:

• Pattern: {prefix}-{project_name} (e.g., a0-myproject)
• Default dataset: a0-default when no project is active
• Change the prefix in plugin settings to create separate knowledge spaces

The plugin includes a companion SKILL.md that can be loaded via skills_tool:load cognee for on-demand CLI access to Cognee operations (search, add data, manage datasets, etc.).

cognee/
├── plugin.yaml                    # Plugin manifest
├── default_config.yaml            # Default settings
├── requirements.txt               # Dependencies (aiohttp)
├── hooks.py                       # install() + pre_update()
├── execute.py                     # Health check / setup verification
├── README.md                      # This file
├── SKILL.md                       # Companion skill for CLI access
├── helpers/
│   ├── __init__.py
│   └── cognee_helper.py           # Core REST API integration (~500 lines)
├── extensions/
│   └── python/
│       ├── monologue_start/
│       │   └── _20_cognee_init.py
│       ├── system_prompt/
│       │   └── _30_cognee_context.py
│       ├── message_loop_prompts_after/
│       │   └── _51_cognee_recall.py
│       └── monologue_end/
│           └── _52_cognee_retain.py
├── prompts/
│   ├── cognee.context.md           # Context injection template
│   ├── cognee.recall.md            # Recall injection template
│   └── cognee.retain_extract.sys.md # LLM extraction system prompt
└── webui/
    └── config.html                 # Settings panel (Alpine.js)

1. Verify COGNEE_BASE_URL is set in Settings → Secrets
2. Check Cognee server is reachable: curl $COGNEE_BASE_URL/api/health
3. Enable Debug Logging in plugin settings to see detailed logs
4. Run the plugin health check from Settings → Plugins → Execute

1. Ensure Retain and Recall are both enabled
2. Check that Auto-Cognify is enabled (data must be cognified before it's searchable)
3. Verify the dataset has data: check Cognee's /api/v1/datasets endpoint
4. Try a different Search Type (e.g., RAG_COMPLETION for simpler retrieval)

1. Reduce Cache TTL to serve more from cache
2. Switch to a lighter search type (CHUNKS or SUMMARIES)
3. Reduce max token limits for recall and context
4. Disable Auto-Cognify and run cognification manually during quiet periods

MIT

• Cognee by Topoteretes — the underlying AI memory platform
• Agent Zero by Jan Tomášek — the agentic framework
• Plugin architecture inspired by a0-hindsight and a0-plugin-honcho
• This plugin is part of the neurocis method for personal assistant memory, powering knowledge retention and semantic search for the Bibiotek agent (librarian)