Memory-Native Event-Driven Architecture for Edge Continual Learning
Important
Read the full MNEMA technical report (PDF)
Architecture, methodology, benchmark results, limitations, and reproducibility in one document.
MNEMA is a NumPy research prototype for learning from a sequence of labeled examples. It combines sparse input codes, a fast associative store, and a slower adaptive spiking cortex to study the trade-offs between accuracy, forgetting, memory allocation, and projected inference cost.
The implementation runs on CPU. It includes six-method Split-MNIST experiments, checkpoint-isolated evaluation, operation counters, and reproducible result artifacts. It does not require pretrained weights or a GPU.
Authors: Ujan Dey and Swapnil · Authorship
Licensed under the Apache License 2.0. Copyright (c) 2026 COGNX.
Full technical report (PDF) · Benchmark report · Get started · Documentation index
| Your goal | Start with |
|---|---|
| Read the complete project report | Full MNEMA technical report (PDF) |
| Understand the model and its limitations | Architecture and data flow |
| Install and verify the research environment | Getting started |
| Run or reproduce an experiment | Benchmark runbook |
| Assess the evidence | Results and provenance |
| Use the model from Python | Python API |
| Make a change | Contributor guide |
Run commands from the repository root using Python 3.11.7. The package metadata and research dependencies now use the same tested environment; see environment setup.
git clone https://github.com/ujandey/mnema.git
cd mnema
python -m venv .venvCreate the environment using Python 3.11.7. Activate it with .venv\Scripts\Activate.ps1 in PowerShell or source .venv/bin/activate in bash, then run:
python -m pip install -r experiments/research_requirements.txt
python -m unittest discover -s tests -p "test_research*.py" -v
python experiments/e0_calibrate.pyThe tests use synthetic data; calibration checks dense operation counts without downloading MNIST. For an end-to-end debug experiment:
python experiments/run_research_benchmark.py --quickThis uses one seed and 100 training / 100 test images per task. The first run downloads and verifies MNIST if needed, then writes to results/research_benchmark/quick/. This checkout already contains a saved quick run: if its source or environment differs, resume will refuse to mix results. See resume and fresh runs before using --force.
Image -> [E] spike encoder -> [S] sparse separator -> [F] fast store --+
[C] slow cortex -+-> [R] prediction
[N] controller gates learning and sleep
[Z] sleep replays store rows into cortex
[X] counters record selected operations
The default separator selects 64 active indices from a 16,384-unit expansion. FastStore accumulates class votes at those addresses; the cortex gathers the corresponding synaptic weights and updates adaptive neuron state. A confidence-dependent readout blends the two outputs. During training, novelty triggers fast writes, prediction errors trigger cortex learning, and accumulated sleep pressure triggers consolidation.
The architecture guide explains the exact implementation, including the collapse of spike timing in the separator and the single-row replay used during sleep.
Three result families exist and use different protocols:
| Result family | Data and replication | Where to read it |
|---|---|---|
| Completed full research benchmark | All 60,000 training / 10,000 test images; seeds 0-9; six methods | Full report and evidence bundle |
| Checked-in debug run | 100 train / 100 test images per task; seed 0; six methods | Debug report |
| Historical experiments | Small subsets; older evaluation and replay behavior | Historical artifact guide |
The full benchmark is complete: 60 of 60 seed-method runs, with all saved-artifact validation checks passing. The report, summaries, plots, raw records, and worker metadata were imported from the supplied cognx_benchmarking folder. The import audit independently checked their consistency without rerunning training or inference. Original files and source identifiers are preserved.
- Energy is projected. Native software counters feed technology-card estimates. Coverage is incomplete, and the results are not measured device energy or an end-to-end training-energy comparison.
- 64 KiB is a FastStore configuration, not total memory. Eviction counts 25 bytes per active row, while the default arrays use 28. The recorded MNEMA resident array payload is 4,391,100 bytes, excluding Python/runtime overhead and temporary workspace.
- Native inference changes state.
is_training=Falseskips learning but still updates encoder, separator, neuron, and controller state. Research evaluation restores the trained checkpoint after each image. - The prototype has a shallow spiking cortex. It is not a deployed neuromorphic system, a multilayer SNN, or evidence of category-level one-shot learning. Sparse codes do not establish a privacy guarantee.
The comparison below uses the completed full run. Accuracy and forgetting are mean +/- sample SD over ten seeds; no statistical-significance claim is made. Memory and inference projections are the reported means. The marked block remains compatible with the report generator; use the full directory when updating it.
| Method | Final ACC (%) | Forgetting (pp) | Resident arrays (bytes) | Projected inference (microjoules/image) |
|---|---|---|---|---|
| Naive MLP | 19.79 +/- 0.06 | 99.52 +/- 0.12 | 814,120 | 1.768 |
| Replay-300 (Research) | 82.44 +/- 1.33 | 20.58 +/- 1.62 | 1,049,636 | 1.768 |
| Replay-64KiB | 67.59 +/- 2.07 | 39.30 +/- 2.60 | 879,291 | 1.768 |
| EWC | 19.98 +/- 0.36 | 99.12 +/- 0.40 | 7,327,080 | 1.768 |
| DER++-300 | 89.39 +/- 0.60 | 12.12 +/- 0.73 | 1,061,636 | 1.768 |
| MNEMA | 77.12 +/- 0.87 | 6.53 +/- 1.41 | 4,391,100 | 0.316 |
MNEMA has the lowest observed mean forgetting; DER++-300 has the highest observed mean accuracy. Replay-300 and DER++-300 both exceed MNEMA's accuracy with less resident array payload. The smaller MNEMA inference projection is partial and is not measured hardware energy.
Detailed report · Summary JSON · Summary CSV · Validation
The full record was created on September 11, 2026, with source commit fe86ec1c6abc600dda8ec50565a551af4e5434bd. The current checkout is not byte-identical to that source manifest; see provenance and reproduction limits. Quick results remain available for pipeline debugging and do not replace this full comparison.
The completed report can be read without running an experiment. To produce a new full run after verifying the research environment, run all 60 seed-method pairs sequentially:
python experiments/run_research_benchmark.py --mode full --seeds 10The included full directory contains the supplied evidence, so a mismatched new run will refuse to resume it. Follow the runbook to preserve or deliberately archive that bundle before starting a fresh run. Completed pairs can be resumed only with matching configuration; an interrupted pair restarts from the beginning.
A lightweight CI workflow runs the research tests and calibration check on pushes and pull requests. The full benchmark workflows remain manual because they are computationally expensive; see the cloud execution guide.
| Path | Purpose |
|---|---|
| mnema/ | Encoder, separator, fast store, cortex, readout, controller, consolidation |
| baselines/ | Historical MLP and research Naive, Replay, EWC, DER++ implementations |
| benchmarks/ | MNIST loading, verification, and task streams |
| instrument/ | Operation counters and technology cards |
| experiments/ | Runners, auditing, validation, plotting, reporting |
| tests/ | Research protocol and distributed artifact integrity checks |
| results/ | Debug and historical artifacts |
| docs/mnema.pdf | Full MNEMA technical report |
| docs/ | Setup, architecture, API, protocol, and troubleshooting |
Use the root main.py for historical benchmarks and plotting. Use the scripts under experiments/ for the reproducible research protocol.
MNEMA's original source code, documentation, and benchmark artifacts are licensed under the Apache License, Version 2.0. Third-party dependencies and the MNIST dataset retain their respective licenses and terms; the project license does not replace them. See CONTRIBUTING.md for development guidance.