File size: 1,991 Bytes
7873200
 
 
 
 
 
 
 
 
 
6a700f4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35f4545
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
---
dataset_info:
  features:
    - name: scenario_id
      dtype: string
  config_name: default
  version: 0.2.0
pretty_name: Mnemosyne Memory Lifecycle Benchmark
---

# Mnemosyne Memory Lifecycle Benchmark Dataset v0.2.0

This deterministic synthetic benchmark contains 40 scenarios across eight
categories and five documented reference strategies. Strategy execution is
blind to expected labels; execution and evaluation are separate phases.

Current results are fixture results, not real-world memory accuracy. Simple
strategies remain competitive on straightforward current-state questions, while
typed lineage uniquely reconstructs the historical point in the current
fixtures. Twenty-five of 40 scenarios are currently non-discriminating and are
preserved rather than hidden. Metrics are `not_applicable` when no valid
denominator exists. Logical erasure is not physical deletion, and this
benchmark does not establish superiority over production memory systems.

Public links:

- Space: https://huggingface.co/spaces/solsticestudioai/mnemosyne-memory-lab
- Direct app: https://solsticestudioai-mnemosyne-memory-lab.static.hf.space/
- Collection: https://huggingface.co/collections/solsticestudioai/solstice-agent-reliability-lab
- Methodology: docs/METHODOLOGY.md
- Scenario design: docs/SCENARIO_DESIGN.md
- Metric validity: docs/METRIC_VALIDITY.md
- Limitations: docs/LIMITATIONS.md
- Validity summary: reports/benchmark-validity-summary.md
- Claim evidence: reports/claim-evidence-matrix.md

## Test Your Own Memory System

Tooling release v0.2.1 provides a provider-neutral, label-free subprocess
protocol. Clone or download the package, copy an adapter template, implement
JSON stdin/stdout, execute against the frozen scenarios, evaluate separately,
build a result bundle, and load it into the Space locally. No result is uploaded
automatically; there is no public leaderboard; users control whether they share
anything. External outputs are community-provided and unverified.