memarena v0.1.2

License compliance release. No code or result changes.

• Include LongMemEval MIT notice with the bundled demo data
  (src/memarena/data/demo/LICENSE-LongMemEval.txt, copied unmodified from xiaowu0162/LongMemEval) and cite it
  in a "Third-party data" section of the README.

Install: uvx --from git+https://github.com/diegoperezg7/memarena@v0.1.2 memarena demo (PyPI coming).

memarena v0.1.1

Polish release after review of 0.1.0. No change to the benchmark logic or the recorded results.

• Terminal: memarena demo and run fit narrow terminals (80 columns, CI logs, pipes). Below 150 columns the
  per-category accuracy moves to a second table; below 100 columns p50 and ingestion time are dropped.
• Viewer: the judge audit is stored as structured data and shown in the viewer's language (EN/ES).
• Viewer: system cards share one layout (kind chip under the name), so they line up.
• Viewer: answers render the bold/italic/code markdown models emit instead of showing raw asterisks; latency and
  token counts sit on their own line in every answer card; the question field is labelled "Question"; Spanish
  cards say "correctas".
• Viewer: deep links #questions, #method and #q=<question id> open a tab or expand a question.

Install: uvx --from git+https://github.com/diegoperezg7/memarena@v0.1.1 memarena demo (PyPI coming).

memarena v0.1.0

First release of memarena: benchmark memory layers on your own conversations, not on someone else's leaderboard.

Highlights

• Import ChatGPT and Claude exports (conversations.json or the .zip) and generic JSONL.
• memarena generate: questions with gold answers and evidence pointers in LongMemEval's five categories, written as pending and reviewed in the viewer before use.
• memarena run: mem0 (OSS), Letta (REST API) and Supermemory (API) always next to three baselines: full context, BM25 and embeddings. Same answer model and prompt for every system; the judge prompt is published in docs/judge-prompt.md.
• Metrics: judge accuracy overall and per category, retrieval latency p50/p95, ingestion time, prompt tokens and cost.
• React viewer bundled in the wheel (memarena view, static export for Pages), ES/EN, light/dark, mobile.
• --max-cost budget checked before every paid call; privacy warning before sending data to a cloud provider. No telemetry.

Recorded demo

memarena demo shows a real run on 20 LongMemEval questions (MIT) over one ~75k-token history: mem0 75%, BM25 70%, embeddings 55%, full context 50% (judge audited by hand: 3 false positives, order unchanged). Example viewer: https://diegoperezg7.github.io/memarena/

Known limits

• Letta and Supermemory adapters are mock-tested only in this release (no API keys available; Letta's self-hosted REST server is end-of-life).
• 20 questions is a small sample.

Install: uvx --from git+https://github.com/diegoperezg7/memarena@v0.1.0 memarena demo (PyPI coming).