Add research decision memory benchmark

examples/benchmarks/research-decision-memory/README.md

@@ -0,0 +1,53 @@
+# Research Decision Memory Benchmark
+
+This benchmark evaluates long-running research and product agents that need to
+remember the current decision trail, cite the evidence that changed a decision,
+and suppress stale assumptions from earlier sessions.
+
+The scenario models a product-research agent over several planning sessions.
+Early assumptions about target users, pricing, launch readiness, data handling,
+and model routing are later corrected by new evidence. The benchmark asks
+golden probes about the final state and scores whether each memory strategy
+returns current facts without reintroducing superseded decisions or synthetic
+secrets.
+
+## Strategies
+
+- `active_decision_digest`: Memanto-style typed digest that stores current
+  decisions by slot, keeps evidence IDs, and redacts erased facts.
+- `passive_graph_history`: graph-style historical state memory that keeps all
+  known values unless a caller reconciles conflicts.
+- `append_only_log`: raw transcript retrieval over every historical note.
+- `recent_window_log`: sliding recent-window retrieval that avoids older stale
+  notes but forgets older still-valid decisions.
+
+## Metrics
+
+- Current-fact accuracy against the golden decision state.
+- Evidence coverage for required source IDs.
+- Stale conflict rate when superseded assumptions appear in the answer.
+- Synthetic secret leak rate.
+- Retrieved token footprint.
+- p95 retrieval latency in milliseconds for live stdout runs.
+- Signal/noise ratio over retrieved evidence.
+
+## Run
+
+No third-party dependencies are required.
+
+```bash
+python examples/benchmarks/research-decision-memory/run_benchmark.py
+python examples/benchmarks/research-decision-memory/run_benchmark.py --output examples/benchmarks/research-decision-memory/results/sample_results.json --markdown examples/benchmarks/research-decision-memory/results/sample_results.md
+python -m unittest discover -s examples/benchmarks/research-decision-memory -p "test_*.py"
+```
+
+Saved JSON and Markdown outputs omit runtime-specific latency by default so
+committed sample artifacts are reproducible. Pass `--include-latency-output`
+when you need to save live p95 latency values.
+
+## Expected Result
+
+The active digest should retain current decisions with evidence, avoid stale
+assumptions, and retrieve less context than the append-only and graph-history
+baselines. The recent-window baseline should avoid many stale facts but miss
+older durable decisions that are still current.

examples/benchmarks/research-decision-memory/requirements.txt

@@ -0,0 +1 @@
+# This benchmark uses only the Python standard library.