Orbit

Memory infrastructure for developer-facing AI applications.

Orbit gives your app long-term memory with adaptive personalization, relevance ranking, and feedback learning.
You integrate once, then use the same simple loop everywhere: ingest -> retrieve -> feedback.

Project status: Alpha (0.1.x)

What Orbit Is

Orbit is a memory layer for AI products where context quality matters over time.

Orbit handles:

• event ingestion and durable memory storage
• semantic retrieval with ranking and intent-aware reweighting
• adaptive inferred memories (learning patterns, progress, preferences)
• feedback-driven ranking updates
• provenance metadata for debugging why a memory was surfaced
• API runtime concerns (JWT auth, rate limits, idempotency, observability)

Why Orbit

Without a memory layer, many assistants either forget important user signals or overload prompts with noisy history.
Orbit is built to return small, high-signal context sets that improve answer quality while staying explainable.

Repository Layout

Path	Purpose
src/orbit/	Public Python SDK (MemoryEngine, AsyncMemoryEngine)
src/orbit_api/	FastAPI runtime (/v1/ingest, /v1/retrieve, /v1/feedback, etc.)
src/memory_engine/	Memory orchestration and adaptive personalization
src/decision_engine/	Core decision/ranking/storage primitives
examples/	Working integration samples
docs/	Deployment, integration, and operational docs

Quickstart (5 Minutes)

1) Install

Published package:

pip install orbit-memory

Local repository install:

pip install -e .

2) Start Orbit API Locally

Recommended full local stack (API + Postgres + Prometheus + OpenTelemetry collector):

docker compose up --build

3) Generate a Local JWT

python scripts/generate_jwt.py \
  --secret orbit-dev-secret-change-me \
  --issuer orbit \
  --audience orbit-api \
  --subject local-dev

4) Integrate with the SDK

from orbit import MemoryEngine

engine = MemoryEngine(
    api_key="<jwt-token>",
    base_url="http://localhost:8000",
)

engine.ingest(
    content="I keep confusing for-loops and while-loops.",
    event_type="user_question",
    entity_id="alice",
)

retrieval = engine.retrieve(
    query="What should I know about alice before answering?",
    entity_id="alice",
    limit=5,
)

if retrieval.memories:
    engine.feedback(
        memory_id=retrieval.memories[0].memory_id,
        helpful=True,
        outcome_value=1.0,
    )

5) Optional Direct API Call

curl -X POST http://localhost:8000/v1/ingest \
  -H "Authorization: Bearer <jwt-token>" \
  -H "Content-Type: application/json" \
  -d '{"content":"User completed first lesson","event_type":"learning_progress","entity_id":"alice"}'

Integration Paths

Mode	Best for	Entry point
Python SDK	Python apps wanting fastest integration	from orbit import MemoryEngine
REST API	Non-Python or service-to-service integration	POST /v1/ingest, GET /v1/retrieve, POST /v1/feedback
OpenClaw plugin	Agent workflows in OpenClaw	integrations/openclaw-memory/

Core Concepts

Concept	Description
entity_id	Stable identity key for a user, agent, or account
ingest	Add a memory signal (user_question, assistant_response, etc.)
retrieve	Fetch ranked context memories for a query
feedback	Send outcome quality signal (helpful, outcome_value)
inferred memory	Auto-generated memory from repeated patterns/feedback
inference provenance	why/when/type/derived_from metadata for traceability

Architecture Overview

1. Application sends events to Orbit.
2. Orbit processes and stores memories.
3. Retrieval ranks candidate memories for each query.
4. Feedback updates ranking behavior and personalization state.
5. Adaptive engine emits inferred memories when confidence thresholds are met.

High-level flow:

App -> Orbit API/SDK -> Memory Engine -> Storage/Ranker -> Retrieved Context -> App

Configuration

Core settings live in .env.example.

Most important environment variables:

Area	Variables
Storage	MDE_DATABASE_URL, MDE_SQLITE_PATH, MDE_EMBEDDING_DIM
Providers	MDE_EMBEDDING_PROVIDER, MDE_SEMANTIC_PROVIDER
Auth	ORBIT_JWT_SECRET, ORBIT_JWT_ISSUER, ORBIT_JWT_AUDIENCE, ORBIT_JWT_REQUIRED_SCOPE
API runtime	ORBIT_API_HOST, ORBIT_API_PORT, ORBIT_ENV
Limits	ORBIT_RATE_LIMIT_EVENTS_PER_DAY, ORBIT_RATE_LIMIT_QUERIES_PER_DAY, ORBIT_RATE_LIMIT_PER_MINUTE
Personalization	MDE_ENABLE_ADAPTIVE_PERSONALIZATION and MDE_PERSONALIZATION_*
Observability	ORBIT_OTEL_SERVICE_NAME, ORBIT_OTEL_EXPORTER_ENDPOINT

Provider Adapters

Supported adapter families:

• OpenAI
• Anthropic
• Gemini
• Ollama
• deterministic/context local fallback

Optional extras:

pip install "orbit-memory[anthropic]"
pip install "orbit-memory[gemini]"
pip install "orbit-memory[ollama]"
pip install "orbit-memory[llm-adapters]"

Production Readiness Checklist

• use PostgreSQL (MDE_DATABASE_URL) instead of SQLite fallback
• run schema migrations (python -m alembic upgrade head)
• enforce strict JWT verification (ORBIT_ENV=production + non-default secrets)
• configure API limits and batch caps
• enable metrics + tracing and scrape /v1/metrics
• set backup/retention policies for primary database
• load-test retrieval quality under long horizons

Monitoring and Operations

Operational endpoints:

• GET /v1/health
• GET /v1/status
• GET /v1/metrics
• GET /v1/memories (paginated memory inspection)

Response headers include:

• X-RateLimit-Limit
• X-RateLimit-Remaining
• X-RateLimit-Reset
• Retry-After (429)
• X-Idempotency-Replayed (write endpoints)

Examples

• Live Orbit + Ollama chatbot: examples/live_chatbot_ollama/
• OpenClaw memory plugin scaffold: integrations/openclaw-memory/
• API + deployment runbook: docs/DEPLOY_RENDER_VERCEL.md

Evaluation and Quality

Baseline scorecard:

python scripts/run_orbit_eval.py --output-dir eval_reports/latest

Long-horizon personalization soak:

python scripts/soak_personalization.py \
  --output-dir soak_reports/latest \
  --sqlite-path tmp/orbit_soak.db \
  --turns-per-persona 500

Troubleshooting

Symptom	Likely cause	Fix
401 Unauthorized	invalid/missing JWT or claim mismatch	verify iss, aud, exp, signing secret
429 Too Many Requests	rate limit exhausted	respect Retry-After, tune limits
409 Conflict on writes	idempotency key reused with different payload	use unique keys per payload
weak personalization	unstable entity_id or missing feedback	keep stable entity mapping and send feedback

Contributing

Local validation commands:

python -m ruff check src tests
python -m mypy src
python -m pytest -q
python -m pylint src tests scripts --fail-under=9.0

Make targets:

make lint
make type-check
make format-check
make test
make migrate
make run-api

Documentation

• Canonical developer integration guide: docs/developer_documentation.md
• Build and implementation log: BUILD_AND_SPEC.md
• Repo metadata for discovery: metaData.md

License, Security, Support

• License: add a LICENSE file before public OSS distribution.
• Security policy: add a SECURITY.md for vulnerability reporting.
• Support: use GitHub Issues for bugs, feature requests, and integration questions.