[codex] Add incident runbook memory benchmark

examples/benchmarks/incident-runbook-memory/README.md

@@ -0,0 +1,77 @@
+# Incident Runbook Memory Benchmark
+
+This benchmark evaluates the core 2026 agent-memory tradeoff from issue #639:
+current-state accuracy versus retrieved-context footprint.
+
+The scenario is an incident-response assistant that receives dense, shifting
+technical notes over five sessions. Facts mutate over time: owners rotate,
+runbooks are superseded, customer-facing language changes, and one raw note
+contains a synthetic credential. The benchmark asks current-state questions and
+scores each backend against golden answers.
+
+## Backends
+
+| Backend | Role |
+|---|---|
+| `active_incident_digest` | Memanto-style active memory that keeps the latest fact per subject/key and redacts secrets before retrieval. |
+| `append_only_log` | Archive-memory control that returns every raw event for the queried subject. |
+| `recent_window_log` | Short-context control that keeps only the five newest raw events. |
+
+The two controls are deliberately simple and deterministic. They are not
+published vendor scores for Mem0, Letta, Zep/Graphiti, or Hindsight. They are
+included so the experimental harness, dataset, metrics, and expected behavior
+can be reviewed without API keys. The same `MemoryBackend` protocol can be used
+to add hosted provider adapters while preserving the dataset and evaluator.
+
+## Metrics
+
+The runner reports:
+
+- retrieval accuracy against golden current facts
+- total ingested tokens
+- stored tokens
+- average retrieved tokens
+- deterministic p95 latency estimate
+- stale conflict rate
+- secret leak rate
+
+Latency is modeled from scanned records and retrieved tokens so CI runs are
+stable. Hosted adapters should replace that value with measured wall-clock
+latency while keeping the same query set.
+
+## Run
+
+No runtime dependency is required beyond Python 3.10+.
+
+```bash
+python run_benchmark.py --output results/sample_results.json --markdown results/sample_results.md
+```
+
+From the repository root:
+
+```bash
+python examples/benchmarks/incident-runbook-memory/run_benchmark.py \
+  --output examples/benchmarks/incident-runbook-memory/results/sample_results.json \
+  --markdown examples/benchmarks/incident-runbook-memory/results/sample_results.md
+```
+
+Run the tests:
+
+```bash
+python -m unittest discover -s examples/benchmarks/incident-runbook-memory -p test_*.py
+```
+
+## Dataset Design
+
+All backends ingest the same ordered event stream. The golden queries cover:
+
+- a service owner that changed from `payments-oncall` to `checkout-platform-oncall`
+- a checkout runbook that superseded `restart-all-pods`
+- customer language that changed for `payments-ledger`
+- a mitigation that changed for `catalog-worker`
+- a failover region that changed from `us-east-1` to `eu-west-1`
+- an early but still-current `billing-cron` owner
+- a synthetic credential that must not be surfaced
+
+This makes the benchmark stress both sides of the challenge: stale-context
+avoidance and resource footprint.

examples/benchmarks/incident-runbook-memory/requirements.txt

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