memory-harness 1.2.6

Benchmark and validate AI memory systems

Memory Harness

Benchmark and validate AI memory systems. Detect regressions, leakage, and shortcuts.

==================================================
 MEMORY BENCHMARK REPORT
==================================================
 Accuracy@1:      83.3%  pass
 Accuracy@3:     100.0%  pass
 Cross-tenant:     0.0%  pass
 Collision:       16.7%  pass
 Confidence:      83.3%  pass
==================================================
 SCORE: 90.0/100  GRADE: A
==================================================

Install

pip install memory-harness httpx

Quick Start

1. Create your dataset (data.jsonl)

{"type":"store","item_id":"doc1","tenant_id":"acme","text":"Customer bought 3 widgets"}
{"type":"store","item_id":"doc2","tenant_id":"acme","text":"Support ticket: login issue"}
{"type":"store","item_id":"doc3","tenant_id":"globex","text":"New user signup from France"}
{"type":"store","item_id":"doc4","tenant_id":"globex","text":"User upgraded to premium"}
{"type":"query","query_id":"q1","tenant_id":"acme","text":"customer purchase","expected_item_id":"doc1"}
{"type":"query","query_id":"q2","tenant_id":"acme","text":"login problem support","expected_item_id":"doc2"}
{"type":"query","query_id":"q3","tenant_id":"globex","text":"new customer france","expected_item_id":"doc3"}
{"type":"query","query_id":"q4","tenant_id":"globex","text":"plan upgrade","expected_item_id":"doc4"}

2. Run benchmark

memorybench dataset -d data.jsonl --provider-endpoint https://your-memory-api.com --n-probe 16

3. Get your score

SCORE: 90.0/100  GRADE: A
PASS (threshold: 70)

Metrics

Metric	What it measures	Target
Accuracy@1	Exact match rate	≥70%
Accuracy@k	Correct item in top-k	≥90%
Cross-tenant	Data leakage between tenants	<5%
Collision	Different queries → same result	<20%
Confidence	Clear winner (margin)	≥80%

Grading

Grade	Score	CI Exit
A	90-100	0 (pass)
B	80-89	0 (pass)
C	70-79	0 (pass)
D	60-69	1 (fail)
F	<60	1 (fail)

CI Integration

Add to .github/workflows/memory-audit.yml:

name: Memory Audit
on: [push, pull_request]

jobs:
  audit:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: '3.11'
      
      - name: Install
        run: pip install memory-harness httpx
      
      - name: Run Memory Benchmark
        run: |
          memorybench dataset \
            -d tests/memory_data.jsonl \
            --provider-endpoint ${{ secrets.MEMORY_API_URL }} \
            --n-probe 16 \
            --pass-threshold 70
      
      - name: Upload Report
        uses: actions/upload-artifact@v4
        if: always()
        with:
          name: memory-report
          path: dataset_report.*

Setup:

1. Add MEMORY_API_URL to repository secrets
2. Create tests/memory_data.jsonl with your test data
3. Push — CI fails if score < 70

Dataset Format

Store items (what to remember):

{"type":"store","item_id":"unique_id","tenant_id":"namespace","text":"content"}

Query items (retrieval tests):

{"type":"query","query_id":"q1","tenant_id":"namespace","text":"search query","expected_item_id":"unique_id"}

Validate before running

memorybench validate -d data.jsonl

CLI Reference

memorybench --version                    # Version
memorybench validate -d FILE             # Validate dataset
memorybench dataset -d FILE [OPTIONS]    # Run benchmark

Options

Flag	Default	Description
-d, --dataset	required	JSONL file
--provider-endpoint	-	Memory API URL
--n-probe	16	Pattern dimension
--pass-threshold	70	Minimum score
-a, --adapter	text	hash, text, embedding
-o, --output	dataset_report.json	Report file

Provider API

Your memory endpoint must implement:

POST /reset   {"seed": int}
POST /store   {"pattern": [[float]], "cue": [[float]], "learn_steps": int}
POST /recall  {"cue": [[float]], "steps": int} → {"pattern": [[float]]}

License

MIT