Agentic Context Engineering Toolkit

Research-oriented framework for Agentic Context Engineering. It captures, ranks, and reuses "context deltas" from LLM interactions so agents adapt without retraining, following the methodology described in Agentic Context Engineering Framework.

Features

• LLM provider agnostic (OpenAI, Anthropic, LiteLLM, Ollama, custom wrappers)
• Storage backend agnostic (memory, SQLite, Postgres/pgvector, extensible interfaces)
• Token budget management, retrieval & ranking, reflection, and curation pipelines
• Ready for Python 3.12 with strict typing, async workflows, and modern tooling

Getting Started

python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate
pip install -r requirements.txt

Project Layout

.
  acet/               # Library source (packages added per phase)
  benchmarks/         # Performance and benchmark suites
  docs/               # Documentation site sources
  examples/           # Usage examples and sample apps
  tests/              # Unit, integration, and benchmark tests

Development Workflow

1. Create/activate the local virtual environment.
2. Install dependencies with pip install -r requirements.txt.
3. Run format and lint checks: black . and ruff check.
4. Run type checks: mypy --strict ..
5. Run tests: pytest --cov=acet.

Performance Snapshot

• Delta retrieval (250 active deltas): ~2 ms mean latency (tests/benchmarks/test_delta_retrieval.py)
• SQLite save/query (300 staged deltas): ~23 ms mean latency (tests/benchmarks/test_storage_throughput.py)
• Curator dedup (300 proposed insights, 30% duplicates): ~140 ms mean latency (tests/benchmarks/test_curator_throughput.py)

All benchmarks are reproducible via the CLI harnesses under benchmarks/. For example:

python benchmarks/delta_retrieval.py --iterations 30 --plot benchmarks/artifacts/delta_latency.png
python benchmarks/storage_throughput.py --backend all --iterations 30 --plot benchmarks/artifacts/storage_latency.png
python benchmarks/curator_throughput.py --proposals 300 --duplicate-ratio 0.3 --iterations 20 --plot benchmarks/artifacts/curator_latency.png

Adjust the parameters or swap in your production embeddings/backends to profile your deployment.