Context-Aware AI Development System

A hybrid approach combining session memory, long-term persistent knowledge, and Claude's advanced tool use features.

Architecture

Tier 1: Session Memory (Hot Context)

• Fast access to git history
• Recent commits and changes
• Current branch state
• Active file tracking

Tier 2: Long-term Memory (Persistent Knowledge)

• Code repositories with semantic search
• Architectural Decision Records (ADRs)
• Production incident history
• Meeting transcripts
• Design artifacts

Tier 3: Intelligent Orchestration

• Tool Search Tool: On-demand tool discovery (85% token reduction)
• Programmatic Tool Calling: Efficient multi-step workflows (37% token reduction)

📖 Deep Dive: See Multi-Project Architecture for how session and long-term memory work together across multiple projects.

Quick Start

1. Clone and Install

# Clone to a standard location (system-wide MCP servers)
git clone https://github.com/your-org/context-aware-ai-system.git ~/.mcp-servers/context-aware-ai-system
cd ~/.mcp-servers/context-aware-ai-system

1. Run the Installer

cd /path/to/luminescent-cluster
./install.sh

# Or with debug logging enabled:
./install.sh --debug

This will:

• Install Python dependencies
• Configure MCP servers in Claude Code (user scope)
• Initialize Pixeltable knowledge base
• Make servers available across all your projects

Options:

• --debug: Enable debug logging (logs to ~/.mcp-servers/logs/pixeltable-memory.log)
• --help: Show usage information

2. Restart Claude Code

Restart Claude Code to load the MCP servers.

3. Ingest Your Codebase

You can now ask Claude Code to ingest your projects directly:

"Ingest this codebase as 'auth-service'"

Or use the MCP tool explicitly:

# Claude will call this tool for you
ingest_codebase(
    repo_path=os.getcwd(),
    service_name="auth-service"
)

Debug Logging

For troubleshooting, enable detailed logging by editing your MCP configuration:

Step 1: Find your Claude Code MCP config:

~/.config/claude/config.json

Step 2: Add env to the pixeltable-memory server:

{
  "mcpServers": {
    "pixeltable-memory": {
      "command": "luminescent-cluster",
      "args": ["pixeltable"],
      "env": {
        "PIXELTABLE_MCP_DEBUG": "1"
      }
    }
  }
}

Step 3: Restart Claude Code

Logs: ~/.mcp-servers/logs/pixeltable-memory.log

Shows:

• Exact paths being ingested
• File counts and timing
• Error details and stack traces
• MCP call parameters

Disable: Remove the "env" field and restart.

See docs/DEBUG_LOGGING.md for details.

4. Add Historical Context

"Add this ADR to the knowledge base" "Record this incident: Auth service outage on Nov 27 due to timeout"

Or use the tools:

# Add an ADR
ingest_architectural_decision(
    adr_path="docs/adr/001-database.md",
    title="ADR 001: Database Choice"
)

# Add an incident
ingest_incident(
    title="Auth Service Outage",
    description="Timeout in auth service caused 500 errors...",
    service="auth-service"
)

File Filtering

Why we filter: The ingestion process filters files by extension to avoid:

• Binary files: Images, executables, compiled artifacts (not useful for text search)
• Generated code: Build outputs, node_modules, vendor directories (adds noise)
• Non-text formats: Don't benefit from embedding-based semantic search

Smart filtering:

• Respects .gitignore: If your project has a .gitignore, it's automatically used to skip files
• Fallback filters: Without .gitignore, skips common patterns (node_modules, pycache, .git, dist, build, etc.)
• Extension filtering: Only ingests source code files (configurable, see below)

Default extensions: Python, JavaScript, TypeScript, Rust, Go, Java, C/C++, Shell, SQL, YAML, Markdown, and more.

Customize for your project:

"Ingest this codebase as 'my-service' with only Rust files"

Or specify custom extensions:

ingest_codebase(
    repo_path=".",
    service_name="my-service",
    extensions=[".rs", ".toml", ".proto"]
)

Check what was ingested:

./scripts/check-status.sh -v  # Shows recent entries with file paths

5. Manage Your Knowledge Base

"Show me stats about the knowledge base" "Create a snapshot called 'pre-release'" "List all services"

MCP Server Configuration

Luminescent Cluster provides two MCP servers that connect to Claude Code (or any MCP-compatible client):

Server	Install	Dependencies
session-memory	Base install	Lightweight (~36 packages)
pixeltable-memory	[pixeltable] extra	Heavy (~500MB macOS, ~2GB Linux/CUDA: torch, sentence-transformers, embedded PostgreSQL)

Most users only need session-memory. The Pixeltable server is for long-term organizational knowledge (ADRs, code embeddings, incident history).

Install the Package

# Recommended: full install with Pixeltable long-term memory (~500MB)
uv tool install "luminescent-cluster[pixeltable]"

# Or use the install script (includes Pixeltable by default)
./install.sh

# Lightweight alternative (session memory only, no ML deps)
uv tool install luminescent-cluster
./install.sh --session-only

Configure Claude Code

Create a .mcp.json in your project root (or add to an existing one). Do not commit this file — it is already in .gitignore.

If the project already has a .mcp.json with other MCP servers, merge the session-memory entry into the existing mcpServers object rather than replacing the file.

Full install (session + Pixeltable — the default):

{
  "mcpServers": {
    "session-memory": {
      "command": "luminescent-cluster",
      "args": ["session"]
    },
    "pixeltable-memory": {
      "command": "luminescent-cluster",
      "args": ["pixeltable"]
    }
  }
}

Development install (editable install in a local venv):

{
  "mcpServers": {
    "session-memory": {
      "command": "/absolute/path/to/.venv/bin/luminescent-cluster",
      "args": ["session"]
    }
  }
}

Use the absolute path to the venv binary because Claude Code does not activate virtual environments when spawning MCP server processes.

Session memory only (if installed with --session-only or without [pixeltable]):

{
  "mcpServers": {
    "session-memory": {
      "command": "luminescent-cluster",
      "args": ["session"]
    }
  }
}

Install Skills

After configuring the servers, install the bundled session management skills:

luminescent-cluster install-skills

This copies /session-init and /session-save skills to .claude/skills/ where Claude Code discovers them as slash commands.

Verify

Restart Claude Code, then run /mcp to confirm the server(s) are connected.

Example Queries

With Claude Code

Once configured, Claude can query both memory tiers:

Session queries (fast):

"What files were changed in the last 24 hours?"
"Show me recent commits about authentication"
"What's the current branch status?"

Long-term queries (semantic):

"What architectural decisions did we make about caching?"
"Have we had any incidents related to database connections?"
"Find all code related to user authentication"

Project-specific queries (with filtering):

"Show me ADRs about database design in the auth-service project"
"What incidents has payment-api had?"
"Search for rate limiting code in the web-app service"

The Pixeltable tools support optional service parameter to filter results to specific projects.

Complex orchestration (programmatic):

"Compare our current auth implementation against the ADR and 
any related incidents to suggest improvements"

This last query would use Programmatic Tool Calling to:

1. Search ADRs for "authentication"
2. Search incidents for "auth-related failures"
3. Search code for "authentication implementation"
4. Synthesize findings without polluting context

Tool Configuration

Tool Search Tool

Available via Claude Code's MCP integration:

• Session memory: Always loaded (defer_loading: false)
• Pixeltable memory: Loaded on-demand (defer_loading: true)
• Additional MCP servers: Defer by default

Programmatic Tool Calling

Enabled via programmaticToolCalling.enabled: true

Allows Claude to write Python orchestration code that:

• Calls multiple tools in parallel
• Processes results in sandbox
• Returns only synthesized output
• Reduces context consumption by 37%

File Structure

luminescent-cluster/
├── pyproject.toml                 # Package config, dependencies, extras
├── src/luminescent_cluster/       # Core package
│   ├── cli.py                     # CLI entry point
│   ├── servers/
│   │   ├── session_memory.py      # Tier 1: Session memory MCP server
│   │   └── pixeltable.py          # Tier 2: Long-term memory MCP server
│   ├── skills/                    # Bundled skills and loader
│   │   ├── loader.py              # SkillLoader with progressive disclosure
│   │   └── bundled/               # Skills shipped in the wheel
│   ├── memory/                    # Memory system (extraction, MaaS)
│   ├── extensions/                # Protocol-based extension system
│   └── chatbot/                   # Multi-platform chatbot gateway
├── tests/                         # Test suite (pytest)
├── docs/adrs/                     # Architecture Decision Records
├── install.sh                     # Global install script (uv tool)
├── quickstart.sh                  # Developer setup script
└── uninstall.sh                   # Uninstall script

Performance Characteristics

Session Memory

• Latency: <10ms (in-memory)
• Scope: Current repository, last 200 commits
• Best for: Hot context, current work

Long-term Memory

• Latency: 100-500ms (semantic search)
• Scope: Entire organizational history
• Best for: Architecture decisions, incident history, cross-service context

Tool Orchestration

• Token savings: 60-85% combined (Tool Search + PTC)
• Accuracy improvement: +7-13% on complex tasks
• Latency improvement: 10x for multi-step workflows

Maintenance

Update Embeddings

Pixeltable automatically updates embeddings when content changes:

# Just update the content, embeddings recompute automatically
kb.update({kb.path == 'some/file.py'}, {'content': new_content})

Create Snapshots

Before major refactors:

from pixeltable_setup import snapshot_knowledge_base

snapshot_knowledge_base(
    name='pre-auth-refactor',
    tags=['v2.0', 'stable']
)

Rollback if Needed

pxt.restore('org_knowledge', snapshot='pre-auth-refactor')

Cost Optimization

Embedding Generation

• Uses local sentence-transformers by default (free)
• Upgrade to OpenAI embeddings if needed

Summaries

• Uses simple truncation by default
• Enable OpenAI summarization in pixeltable_setup.py for better quality

Token Usage

• Tool Search Tool: 85% reduction
• Programmatic Tool Calling: 37% reduction
• Combined effect: ~90% reduction for complex queries

Python Version Requirements

CRITICAL: The Pixeltable database is bound to the Python version that created it. Using a different Python minor version will cause a silent segmentation fault (exit code 139).

Version Compatibility

Created With	Safe to Run	Unsafe
3.10.x	3.10.0 - 3.10.99	3.9.x, 3.11+
3.11.x	3.11.0 - 3.11.99	3.10.x, 3.12+
3.12.x	3.12.0 - 3.12.99	3.11.x, 3.13+

Patch version changes are SAFE (3.11.0 -> 3.11.9). Only minor version changes are dangerous.

Runtime Protection

The MCP servers include a version guard that:

• Creates a .python_version marker on first run
• Exits with code 78 if Python version mismatches
• Exits with code 65 for legacy databases without markers

Quick Fix

# Check what version the database expects
cat ~/.pixeltable/.python_version

# Switch to the correct version
uv venv --python 3.11
source .venv/bin/activate

For migration procedures, see ADR-001.

Troubleshooting

"No git repository found"

Session memory server needs to run in a git repository directory.

"Could not connect to org_knowledge"

Run python pixeltable_setup.py first to initialize the knowledge base.

Tools not appearing in Claude

Check .mcp.json exists and MCP servers are configured correctly.

Exit code 78: Python version mismatch

The runtime guard detected that your Python version doesn't match the database.

# Check expected version
cat ~/.pixeltable/.python_version

# Switch to correct version
uv venv --python <version>
source .venv/bin/activate

Exit code 65: Legacy database detected

The database was created before version tracking was implemented.

# If you know the Python version that created it:
echo '3.11' > ~/.pixeltable/.python_version

# Then run with that version
uv venv --python 3.11
source .venv/bin/activate

Exit code 139: Segmentation fault

Caused by corrupted UDFs (User-Defined Functions) after Python version changes.

Quick diagnosis:

python -m scripts.db_repair --check

Recovery options:

1. Use correct Python version (recommended):

   uv venv --python 3.11  # Use version that created DB
   source .venv/bin/activate

2. Backup and restore (preserves data):

   python -m scripts.backup_restore --backup-restore --confirm

3. Fresh install (deletes all data):

   rm -rf ~/.pixeltable/

See docs/KNOWN_ISSUES.md for detailed recovery procedures.

Architecture: Extension System

Luminescent Cluster uses a Protocol/Registry pattern for extensibility (see ADR-005).

Extension Points

Extension	Purpose	OSS Default
TenantProvider	Multi-tenancy isolation	None (single-user)
UsageTracker	Usage metering/billing	None (no tracking)
AuditLogger	Compliance audit logs	None (local logs only)

Usage Pattern

from src.extensions import ExtensionRegistry

# Check if extensions are registered
registry = ExtensionRegistry.get()

# OSS mode: Extensions are None, code handles gracefully
if registry.tenant_provider:
    tenant_id = registry.tenant_provider.get_tenant_id(context)
    filter = registry.tenant_provider.get_tenant_filter(tenant_id)

# Check mode
registry.get_status()  # {'mode': 'oss', ...} or {'mode': 'cloud', ...}

Implementing Extensions

Extensions implement Python Protocols (duck typing):

from src.extensions import ExtensionRegistry

class MyTenantProvider:
    def get_tenant_id(self, ctx: dict) -> str:
        return ctx.get("x-tenant-id")

    def get_tenant_filter(self, tenant_id: str) -> dict:
        return {"tenant_id": {"$eq": tenant_id}}

    def validate_tenant_access(self, tenant_id, user_id, resource) -> bool:
        return True  # Your RBAC logic

# Register at startup
registry = ExtensionRegistry.get()
registry.tenant_provider = MyTenantProvider()

Chatbot Platform Integrations (ADR-006)

Luminescent Cluster supports conversational interfaces via chatbot integrations on Slack, Discord, Telegram, and WhatsApp. See ADR-006 for full details.

Access Control Configuration

The chatbot uses a pluggable access control system:

from src.chatbot.access_control import (
    DefaultAccessControlPolicy,      # OSS: allow all
    ConfigurableAccessControlPolicy,  # Self-hosted: config-based
    ResponseFilterPolicy,             # Filter sensitive data
)

# Default (OSS mode) - allows all channels and commands
policy = DefaultAccessControlPolicy()

# Self-hosted with restrictions
policy = ConfigurableAccessControlPolicy(
    allowed_channels=["#general", "#engineering"],
    blocked_channels=["#hr", "#legal"],
    allowed_commands=["/help", "/ask", "/search"],
)

# Filter sensitive data in public channels
filter_policy = ResponseFilterPolicy(
    sensitive_patterns=[
        r"password\s*[:=]\s*\S+",
        r"api[_-]?key\s*[:=]\s*\S+",
    ]
)

Context Persistence

Conversation context persists in Pixeltable with 90-day retention:

from src.chatbot.context import ThreadContextManager, PixeltableContextStore

# With persistence
store = PixeltableContextStore()
manager = ThreadContextManager(context_store=store)

# Get/update thread context
context = await manager.get_context(thread_id="thread-123")
await manager.update_context(thread_id="thread-123", message=new_message)

Observability

ChatMetrics provides telemetry for monitoring:

from src.chatbot.metrics import ChatMetrics

metrics = ChatMetrics()
await metrics.record_query(
    platform="discord",
    user_id="user-123",
    query_type="search",
    latency_ms=245,
    tokens_used=150,
    memory_hits=3,
)

Data Management (Self-Hosted)

When using luminescent-cluster self-hosted, you have full control over your data:

• Location: Data stored in your local Pixeltable instance (~/.pixeltable/)
• Deletion: Use Pixeltable CLI or API to manage/delete data
• No Third Party: Amiable does not access your self-hosted data
• Retention: Default 90-day TTL for conversation context (configurable)

For GDPR compliance in self-hosted deployments, you are the data controller. Use Pixeltable's built-in tools:

import pixeltable as pxt

# View stored data
pxt.list_tables()

# Delete conversation context
table = pxt.get_table('conversation_context')
table.delete(table.thread_id == 'thread-to-delete')

# Clear all chatbot data
pxt.drop_table('conversation_context')

Note: For managed Luminescent Cloud deployments, GDPR-compliant /forget-me and /export-my-data commands are available.

Project Structure

luminescent-cluster/
├── pyproject.toml                 # Package config, deps, extras ([pixeltable], [dev], [all])
├── src/luminescent_cluster/
│   ├── cli.py                     # CLI entry point (session, pixeltable, install-skills)
│   ├── version_guard.py           # Python version safety (ADR-001)
│   ├── servers/
│   │   ├── session_memory.py      # Tier 1: Session memory MCP server
│   │   └── pixeltable.py          # Tier 2: Long-term memory MCP server
│   ├── skills/                    # Bundled skills and progressive disclosure loader
│   │   ├── loader.py              # SkillLoader (Level 1/2/3)
│   │   └── bundled/               # Skills shipped in the wheel
│   ├── memory/                    # Memory system (extraction, evaluation, MaaS)
│   ├── extensions/                # Extension system (ADR-005)
│   │   ├── protocols.py           # TenantProvider, UsageTracker, AuditLogger
│   │   └── registry.py            # ExtensionRegistry singleton
│   ├── chatbot/                   # Chatbot platform integrations (ADR-006)
│   │   ├── gateway.py             # Central Chat Gateway
│   │   ├── context.py             # Thread context management
│   │   ├── metrics.py             # ChatMetrics telemetry
│   │   ├── access_control.py      # Access control policies
│   │   └── adapters/              # Platform-specific adapters
│   └── integrations/              # PAT integrations (GitHub, GitLab)
├── tests/
│   ├── test_version_guard.py      # Version guard tests (19 tests)
│   ├── test_extensions.py         # Extension system tests (30 tests)
│   ├── test_skills.py             # Skills loader tests (45 tests)
│   └── chatbot/                   # Chatbot tests
├── docs/adrs/                     # Architectural Decision Records
├── install.sh                     # Global install script (uv tool)
├── quickstart.sh                  # Developer setup script
└── uninstall.sh                   # Uninstall script

Contributing

This is an open-source project under Apache 2.0 license. Contributions welcome!

Note: First-time contributors will need to sign a Contributor License Agreement (CLA). The CLA Assistant will guide you through the process when you open your first PR.

Development Setup

# Clone and quick start
git clone https://github.com/amiable-dev/luminescent-cluster.git
cd luminescent-cluster
./quickstart.sh                    # or: ./quickstart.sh --with-pixeltable

# Or manually
uv venv && uv pip install -e ".[dev]"

# Run tests
pytest tests/ -v --ignore=tests/test_pixeltable_mcp_server.py

Contribution Areas

1. Core Features: MCP server improvements, semantic search enhancements
2. Integrations: Webhook support, IDE plugins, additional Git hosting providers
3. Documentation: Examples, tutorials, API documentation
4. Testing: Additional test coverage, integration tests

Test Suite

# Run all tests
pytest tests/ -v --ignore=tests/test_pixeltable_mcp_server.py

# Run specific test categories
pytest tests/test_skills.py -v              # Skills loader (45 tests)
pytest tests/test_extensions.py -v          # Extension system (30 tests)
pytest tests/test_version_guard.py -v       # Version safety (19 tests)
pytest tests/test_session_memory_mcp_server.py -v  # Session MCP server

ADR Process

Significant changes require an ADR (Architectural Decision Record):

1. Copy docs/adrs/template.md to docs/adrs/ADR-NNN-title.md
2. Fill in context, decision, and consequences
3. Submit PR for review

License

Apache 2.0 License - see LICENSE file

References

• Pixeltable Documentation
• Claude Advanced Tool Use
• MCP Protocol

Advanced Usage & Fallbacks

Advanced Usage & Fallbacks

Enabling Advanced Tool Use (Claude Code)

To leverage Anthropic's Advanced Tool Use features (Tool Search and Programmatic Tool Calling) with this system, you must configure your client (Claude Code).

Configuration (.mcp.json):

{
    "toolConfiguration": {
        "toolSearch": {
            "enabled": true,
            "provider": "regex"  // or "embedding"
        },
        "programmaticToolCalling": {
            "enabled": true
        },
        "deferredLoading": {
            "pixeltableMemory": true  // Defer heavy tools
        }
    }
}

How it works:

• Tool Search: When enabled, Claude Code automatically handles the "beta headers" and tool discovery process. You do not need to implement search_tools yourself; the client handles it.
• Programmatic Tool Calling: Claude Code will write Python orchestration scripts to call our atomic tools (get_recent_commits, search_knowledge) efficiently.

Graceful Fallback (RAG Pattern)

For other AI clients or models that do not support tool calling (or if you prefer manual control), the system fully supports the RAG (Retrieval Augmented Generation) pattern.

How to use:

1. Pre-fetch Context: Use the provided scripts to search for relevant information.
2. Inject Context: Insert the retrieved text into your prompt.

Example (Python):

# 1. Retrieve context programmatically
context = search_knowledge(kb, query="database schema", limit=2)

# 2. Construct prompt
prompt = f"""
Context: {context}
Question: How do I query the user table?
"""

# 3. Send to LLM
response = llm.complete(prompt)

See examples/example_usage.py (Example 9) for a complete working demonstration.