stixdb-engine 0.1.0

StixDB — Reasoning Agentic Context Database. An autonomous AI agent lives inside the DB itself, managing a self-organizing graph memory layer for other agents.

StixDB — Living Memory for AI Agents

Stop shoving documents into filing cabinets. Give your agent an Autonomous Librarian.

StixDB exposes a Search API and an Ask API over your private knowledge —
documents, agent memories, ingested data — with the same interface shape as Perplexity:
ranked results from Search, grounded cited answers from Ask.

Under the hood every collection runs an autonomous agent that continuously
reorganises memory, merges duplicates, decays stale facts, and reasons over a live graph.

---

StixDB = Private Perplexity for Your Data

Perplexity's Sonar API searches the live web and returns cited answers. StixDB does the same for your private knowledge base.

	Perplexity Sonar	StixDB
Data source	Live web	Your documents, memories, agent data
Search API	Web search → ranked results	/search → ranked nodes across collections
Ask API	Web context → cited LLM answer	/ask → graph-grounded cited LLM answer
Citations	Source URLs	Source nodes with content, tier, score
Memory	Stateless	Self-organising graph — decays, merges, promotes
Streaming	✅	✅ /chat/completions with "stream": true
OpenAI-compat	❌	✅ Drop-in at POST /chat/completions
Private / on-prem	❌	✅ Fully self-hosted
Requires API key	Yes (pplx-…)	Search API: No — Ask API: Yes (any LLM provider)
Requires Docker	N/A	No — pip install stixdb-engine is enough to start

The two APIs

Search API — fast ranked retrieval, no LLM:

# Perplexity style: search your private data
POST /search
{
  "query": "What are the upcoming project deadlines?",
  "collections": ["project", "team"],
  "max_results": 5
}

# Response: ranked nodes with scores, tiers, and source lineage

Ask API — grounded, cited answers:

# Ask your private knowledge base
POST /collections/project/ask
{
  "question": "Who is responsible for the payments deadline?",
  "top_k": 10
}

# Response: synthesised answer + citations + confidence + reasoning trace
{
  "answer": "Alice (lead engineer, payments team) owns the June 1st deadline.",
  "confidence": 0.93,
  "sources": [
    { "content": "Alice is the lead engineer on the payments team", "score": 0.96 },
    { "content": "Project deadline is June 1st, 2026", "score": 0.91 }
  ]
}

OpenAI-compatible endpoint — drop in anywhere:

POST /chat/completions
{
  "model": "project",        # collection name becomes the model
  "messages": [{ "role": "user", "content": "What are the upcoming milestones?" }],
  "stream": true
}

Works out of the box with the OpenAI Python SDK, LangChain ChatOpenAI, or any tool that speaks the OpenAI spec — just point base_url at your StixDB instance.

---

Stop Shoving Documents into Filing Cabinets

Standard RAG is a mess of duplicates and stale facts. You shove documents in, retrieve the nearest neighbors, and hope the LLM stitches them into something useful.

Traditional RAG:

  User Question
       │
       ▼
  [Vector Search]  ──→  top-k chunks (static, never changes)
       │
       ▼
  [LLM Prompt]     ──→  answer (no memory of what was useful)

The filing cabinet never organises itself. Stale facts stay forever. Duplicates pile up. Hot information gets no priority over cold. Nothing learns from access patterns.

StixDB is an Autonomous Librarian.

StixDB:

  User Question
       │
       ▼
  [7-Phase Retrieval]  ──→  contextual answer + reasoning trace + cited sources
       │
       ▲
  [Living Memory Graph]
       │
  ┌────┴──────────────────────────────────────────────┐
  │  MemoryAgent (runs every 30 seconds)               │
  │                                                    │
  │  PERCEIVE  ──  track which nodes are accessed      │
  │  PLAN      ──  score heat: 0.6×frequency +         │
  │                            0.4×recency             │
  │  ACT       ──  promote hot nodes → working memory  │
  │                merge semantically similar nodes    │
  │                decay importance by half every 48h  │
  │                prune nodes below 0.05 importance   │
  └────────────────────────────────────────────────────┘

---

What Makes StixDB Different

Capability	Traditional RAG	StixDB
Retrieval	Vector similarity	7-phase: vector → graph BFS → tier-aware re-rank → LLM reason
Memory structure	Flat chunks	Typed graph nodes with edges and clusters
Stale data	Stays forever	Exponential decay (importance × 2^(-t/48h))
Duplicates	Pile up	Auto-merged when cosine similarity > 0.88
Hot vs cold	No concept	5 tiers: working → episodic → semantic → procedural → archived
Access patterns	Ignored	LRU+LFU hybrid heat scoring drives tier promotion
Answers	Raw chunks	LLM synthesis with citations and reasoning trace
Background work	None	Autonomous perceive → plan → act loop per collection
Source lineage	Lost on re-chunk	Preserved across merges (parent_node_ids, char offsets)
LLM dependency	Required	Optional — heuristic mode works with no API key

---

Getting Started

Local development (no Docker):

pip install "stixdb-engine[local-dev]"

See QUICKSTART.md for a complete guide to building agent memory on your laptop.

Production (Docker):

docker compose up -d

See PRODUCTION.md for scaling, multi-agent deployments, and operations.

---

How It Works

Memory Node Types

fact        →  "Project deadline is June 1st, 2026"
entity      →  "Alice — lead engineer, payments team"
event       →  "Sprint review on April 10th"
concept     →  "Payments module uses event-sourcing pattern"
procedure   →  "How to deploy to production"
summary     →  [auto-generated merge of related facts]
question    →  "What are the upcoming deadlines?"  (cached answers)

Memory Tiers

working     ─── hot, frequently accessed  ← +0.15 retrieval boost
episodic    ─── recent, not yet generalised
semantic    ─── generalised knowledge (often from consolidation)
procedural  ─── skills and how-to sequences
archived    ─── cold, eligible for pruning

The agent automatically promotes and demotes nodes based on access heat.

The 7-Phase Retrieval Pipeline

1. Embed Query       →  384-dim vector for the question
2. Vector Search     →  top-15 semantic candidates (threshold: 0.25)
3. Graph BFS         →  expand to neighbours (depth 2)
4. Tier Re-rank      →  working memory nodes get +0.15 score boost
5. Truncate          →  keep top 20 nodes by combined score
6. LLM Reason        →  synthesise answer with citations
7. Record & Trace    →  update access counts, emit telemetry

The Background Agent Cycle (every 30 seconds)

# PERCEIVE
access_data = planner.collect_access_patterns()

# PLAN — hybrid heat score
heat = 0.6 * frequency_score   # saturates at 10 accesses/24h
     + 0.4 * recency_score     # half-life: 12 hours

# ACT
if heat > 0.65:  promote → working memory
if decay < 0.08: demote  → archived
if similarity(node_a, node_b) > 0.88:  merge → summary node
if importance < 0.05 and tier == archived:  prune permanently

---

RAG vs StixDB — Side by Side

Scenario: A month of project updates ingested

RAG after 30 days

results = vector_db.search("project status", top_k=5)
# Returns 5 chunks, including:
# - "Sprint 1 complete" (from day 1 — stale, never removed)
# - "Sprint 1 complete" (duplicate, from a different doc format)
# - "Sprint 2 in progress" (from week 2 — outdated)
# - "Sprint 4 kicked off" (current)
# - "Alice joined the team" (relevant)
#
# No synthesis. No ranking by freshness. No deduplication.
# The LLM now has contradictory context and must figure it out.

StixDB after 30 days

response = await engine.ask("my_agent", "What is the project status?")
# Working memory (tier=working, heat=0.91):
#   → "Sprint 4 kicked off — focus is payments integration"
#   → "Alice is leading sprint 4 delivery"
#
# Auto-merged:
#   → "Sprint 1 complete" + "Sprint 1 done" → single summary node (archived)
#
# Decayed and pruned:
#   → "Sprint 2 in progress" — importance=0.03, pruned at day 22
#
# Response:
#   answer: "Sprint 4 is currently active, focused on payments integration.
#            Alice is leading delivery."
#   sources: [cited working-memory nodes]
#   reasoning_trace: [step-by-step]

---

Architecture

StixDBEngine
├── MemoryGraph          — unified graph + vector interface
│   ├── StorageBackend   — graph topology (NetworkX / KuzuDB / Neo4j)
│   └── VectorStore      — semantic search (NumPy / ChromaDB / Qdrant)
│
├── MemoryAgent          — per-collection autonomous agent
│   ├── AccessPlanner    — LRU+LFU heat scoring → tier promotion
│   ├── Consolidator     — cosine merge (0.88) + decay + prune (0.05)
│   └── MemoryAgentWorker— async background loop (APScheduler)
│
├── ContextBroker        — 7-phase retrieval + LLM reasoning
│   └── Reasoner         — OpenAI / Anthropic / Ollama / Custom / None
│
└── FastAPI Server       — REST API + OpenAI-compatible endpoint

---

Storage & Vector Backends

Graph Storage

Backend	Mode	Use Case	Install
NetworkX	memory	Learning / testing — data lost on exit	Included
KuzuDB	kuzu	Local development — persistent on disk	pip install stixdb-engine[local-dev]
Neo4j	neo4j	Production — scalable with Docker	pip install stixdb-engine[neo4j]

Vector Search

Backend	Scale	Install
NumPy	Up to ~500k nodes	Included
ChromaDB	Medium scale	pip install stixdb-engine (included)
Qdrant	Billion-scale	pip install stixdb-engine[qdrant]

LLM Providers

LLMProvider.OPENAI     # gpt-4o, gpt-4-turbo, ...
LLMProvider.ANTHROPIC  # claude-3-5-sonnet, claude-3-opus, ...
LLMProvider.OLLAMA     # llama3, mistral, phi3, ... (local)
LLMProvider.CUSTOM     # any OpenAI-compatible endpoint
LLMProvider.NONE       # heuristic mode — no API key needed

Embedding Providers

EmbeddingProvider.SENTENCE_TRANSFORMERS  # all-MiniLM-L6-v2, local, free
EmbeddingProvider.OPENAI                 # text-embedding-3-small/large
EmbeddingProvider.OLLAMA                 # nomic-embed-text, local
EmbeddingProvider.CUSTOM                 # any OpenAI-compatible endpoint

---

API & SDK

StixDB exposes:

• REST API: HTTP endpoints for store, retrieve, ask, upload, agent inspection
• OpenAI-compatible endpoint: /v1/chat/completions for drop-in replacement
• Python SDK: stixdb-sdk — sync and async clients

Full API reference and examples:

• See QUICKSTART.md for local development examples
• See PRODUCTION.md for deployment, scaling, and monitoring

---

Resources

• QUICKSTART.md — Build agent memory on your laptop (local development)
• PRODUCTION.md — Deploy with Docker for scaling and multi-agent
• Cookbooks — Runnable examples for all SDK patterns
• Architecture Docs — Deep dives into design and performance
• Contributing — How to contribute to the project

---

Development

Running tests

pip install stixdb-engine[dev]
pytest tests/ -v

Building the SDK

pip install build twine
python -m build

---

Project Layout

stixdb/             Core engine (graph, agent, broker, API)
sdk/                Lightweight Python HTTP client (stixdb-sdk)
examples/           Runnable examples (core, agents, search, OpenAI-compat)
demos/              Standalone sandboxes for experimentation
scripts/            Benchmarks, debug helpers, manual test clients
doc/                Architecture and performance documentation
tests/              Automated test suite

---

One-time setup

1. Create a PyPI account

Register at pypi.org and create an API token: Account settings → API tokens → Add API token (scope: entire account for first upload, then per-project).

2. Configure trusted publishing (recommended — no stored secrets)

In your GitHub repo go to Settings → Environments → New environment and name it pypi.

Then on PyPI go to your project page → Settings → Publishing and add a trusted publisher:

• Owner: your-org
• Repository: stixdb
• Workflow: publish-sdk.yml
• Environment: pypi

This lets GitHub Actions publish without storing an API key anywhere.

3. Install build tools locally

pip install build twine hatch

---

Publishing stixdb-sdk

Manual publish (first time or hotfix)

cd sdk

# Bump the version in sdk/pyproject.toml, then:
python -m build                  # creates dist/stixdb_sdk-x.y.z.tar.gz and .whl
twine check dist/*               # verify the package is valid
twine upload dist/*              # uploads to PyPI — prompts for token

Automated publish via GitHub Actions

# Bump version in sdk/pyproject.toml, commit, then tag:
git tag sdk-v0.2.0
git push origin sdk-v0.2.0
# The publish-sdk.yml workflow triggers automatically

---

Publishing stixdb-engine

# From the repo root:
python -m build                  # creates dist/ from root pyproject.toml
twine check dist/*
twine upload dist/*

Add a publish-engine.yml workflow following the same pattern as publish-sdk.yml (trigger on tags matching engine-v*).

---

Version bump checklist

• Update version in sdk/pyproject.toml (for SDK releases)
• Update version in pyproject.toml (for engine releases)
• Update __version__ in sdk/src/stixdb_sdk/__init__.py
• Update __version__ in stixdb/__init__.py
• Add a CHANGELOG.md entry
• Tag the commit and push

---

Contributing

We welcome contributions of all kinds — bug reports, features, docs, examples.

See CONTRIBUTING.md to get started.

If you find a security issue, see SECURITY.md.

---

License

MIT — see LICENSE.

---

Built with the vision of making AI agents smarter through intelligent, autonomous memory.