Context Memory And Compression

MirrorNeuron uses the Membrane projects for shared working memory, context selection, context compression, and benchmark evaluation.

Components

Python SDK

Install the Membrane Python SDK from source:

cd Membrane/mn-context-engine-python-sdk
python3 -m pip install -e ".[dev]"
python3 -m pytest -q

Optional extras are package-specific:

python3 -m pip install -e ".[compression]"
python3 -m pip install -e ".[qdrant]"

Use qdrant and qdrant-gpu in separate environments because their FastEmbed dependencies are mutually exclusive.

Optimizer Runtime

Install the deterministic optimizer:

cd Membrane/mn-context-auto-optimizer
python3 -m pip install -e ".[dev]"
python3 -m pytest -q

Inspect runtime capabilities:

mn-context-packer runtime-info

Compress a context packet from standard input:

cat packet.json | mn-context-packer compress \
  --compression-mode graph_nlp \
  --target-tokens 800 \
  --focus-id goal_1 \
  --agent-role executor

Supported compression modes:

Benchmarks

Install the benchmark package:

cd Membrane/mn-context-auto-optimizer-benchmark
python3 -m pip install -e ".[dev]"

Run the default graph benchmark:

mn-context-benchmark --config configs/default.yaml

Build a blueprint-derived benchmark suite from the local catalog:

mn-context-build-blueprint-suite \
  --blueprint-root ../../mn-blueprints \
  --packet-output artifacts/data/blueprint_packet_results.json \
  --working-memory-output artifacts/data/blueprint_working_memory_cases.json \
  --coverage-output artifacts/data/blueprint_suite_coverage.json \
  --cases-per-manifest 12

Notes

• The deterministic runtime path should preserve goals, constraints, source references, failures, recovery state, and next actions.
• Optional model or GPU dependencies should be installed only for the benchmark or compression path that needs them.
• Keep private or role-restricted memory out of shared context packets unless the caller explicitly has access.

Related Pages

• Runtime Architecture
• Component Guide
• Security Model