CaspianG · CaspianG · Jul 28, 2026 · Jul 28, 2026 · Jul 28, 2026
diff --git a/README.md b/README.md
@@ -524,10 +524,12 @@ loopback, and production evidence.
 | Direct Memory OS public runs | LoCoMo 1977 queries; LongMemEval-S 470 queries; LongMemEval-V2 Small 451 questions |
 | Strict production evidence | 5/8 requirements |
 | LongMemEval evidence retrieval | WaveMind recall@5 `0.782` |
+| Real LoCoMo memory systems | WaveMind recall@5 `0.548`; Mem0 OSS `0.500`; Hindsight OSS `0.316` |
 | Large-N profiles | 10M Qdrant, 10M sharded Qdrant, 10M pgvector, 50M FAISS |
 
 These results do not claim universal vector-database leadership or completed
 remote 100M/multi-region proof. See the full [Benchmark Guide](docs/BENCHMARKS.md),
+[real public memory-system report](benchmarks/PUBLIC_MEMORY_COMPETITORS.md),
 [living dashboard](https://caspiang.github.io/wavemind/), and
 [Benchmark Brief](docs/BENCHMARK_BRIEF.md) for methods, commands, limitations,
 and machine-readable artifacts.

diff --git a/benchmarks/AGENT_IMPACT.md b/benchmarks/AGENT_IMPACT.md
@@ -1,6 +1,6 @@
 # WaveMind Agent Impact Leaderboard
 
-Generated: `2026-07-28T17:57:37Z`.
+Generated: `2026-07-28T22:43:41Z`.
 
 Agent-impact rows come from checked-in benchmark artifacts. They show behavioral lift on the configured tasks; they do not claim general agent success outside the listed scenarios.
 

diff --git a/benchmarks/BENCHMARK_LEADERBOARD.md b/benchmarks/BENCHMARK_LEADERBOARD.md
@@ -1,7 +1,7 @@
 # WaveMind Benchmark Leaderboard
 
 Generated from `benchmarks/benchmark_matrix_results.json`.
-Last refresh: `2026-07-28T17:57:37Z` from `d011eb848e03`.
+Last refresh: `2026-07-28T22:43:41Z` from `4b6d7dd3e938`.
 
 This is a compact reader-facing view of checked-in benchmark results. It is not a universal vector-database leaderboard: each row uses the primary quality metric for that benchmark, and latency is shown separately so quality wins are not confused with speed wins.
 
@@ -17,7 +17,7 @@ This is a compact reader-facing view of checked-in benchmark results. It is not
 | Long-term memory evidence | long-term-agent-memory | evidence recall@k | WaveMind: 1 / 6.103 ms | Static vector: 1 / 0.648 ms | Quality tie; WaveMind slower |
 | BEIR-style open retrieval runner | retrieval | precision@1 | WaveMind: 0.24 / 117.0 ms | Chroma: 0.243 / 1.794 ms | Baseline leads on quality |
 | [NoMIRACL Russian retrieval](https://huggingface.co/datasets/miracl/nomiracl) | multilingual-retrieval | precision@1 | WaveMind: 0.41 / 10.2 ms | Chroma: 0.41 / 2.603 ms | Quality tie; WaveMind slower |
-| [LoCoMo evidence retrieval runner](https://github.com/snap-research/locomo) | long-term-conversation-memory | evidence recall@k | WaveMind sentence: 0.547 / 3.438 ms | Qdrant sentence: 0.409 / 124.3 ms | WaveMind leads on quality |
+| [LoCoMo evidence retrieval runner](https://github.com/snap-research/locomo) | long-term-conversation-memory | evidence recall@k | WaveMind current public protocol: 0.548 / 4.876 ms | Mem0 OSS: 0.5 / 270.2 ms | WaveMind leads on quality |
 | [LongMemEval evidence retrieval](https://github.com/xiaowu0162/LongMemEval) | long-term-agent-memory | evidence recall@k | WaveMind: 0.782 / 7.274 ms | Static vector: 0.52 / 0.083 ms | WaveMind leads on quality |
 | [LongMemEval evidence 50-query smoke](https://github.com/xiaowu0162/LongMemEval) | long-term-agent-memory | evidence recall@k | WaveMind: 0.92 / 15.3 ms | Static vector: 0.6 / 0.337 ms | WaveMind leads on quality |
 | [ANN index latency curve](https://github.com/erikbern/ann-benchmarks) | index-latency | Recall@k | WaveMind numpy: 1 / 1.988 ms | Qdrant local: 1 / 33.8 ms | Quality tie; WaveMind faster |
@@ -37,7 +37,7 @@ This is a compact reader-facing view of checked-in benchmark results. It is not
 
 | area | current source | claim status | next action |
 |---|---|---|---|
-| Artifact freshness | local matrix refresh at `2026-07-28T17:57:37Z` | source `d011eb848e03`; audit gate enforced by `validate_benchmark_artifacts.py` | Keep weekly refresh green before public claims. |
+| Artifact freshness | local matrix refresh at `2026-07-28T22:43:41Z` | source `4b6d7dd3e938`; audit gate enforced by `validate_benchmark_artifacts.py` | Keep weekly refresh green before public claims. |
 | Serverless telemetry | loopback API pool; `loopback-api-capacity-estimate`; 4 measured replicas | observed SLO `True`; loopback evidence, not a managed-serverless claim | Run `.github/workflows/serverless-observed-telemetry.yml` against deployed API nodes. |
 | External HTTP cluster load | kubernetes-kind-non-loopback-ci; `kubernetes-pod-dns-physical-node-drill`; 4 nodes | SLO `True`; non-loopback Kubernetes pod-DNS evidence | Run `.github/workflows/external-http-cluster-load.yml` with a remote node manifest. |
 | External HTTP active-active loopback | local-loopback; `loopback-api-regions`; 3 regions | SLO `True`; external URL contract over local API regions | Run `.github/workflows/external-http-active-active.yml` with remote regions for production evidence. |

diff --git a/benchmarks/BENCHMARK_REPORT.md b/benchmarks/BENCHMARK_REPORT.md
@@ -1,7 +1,7 @@
 # WaveMind Benchmark Report
 
 This report is generated from `benchmarks/benchmark_matrix_results.json`.
-Last refresh: `2026-07-28T17:57:37Z` from `d011eb848e03`.
+Last refresh: `2026-07-28T22:43:41Z` from `4b6d7dd3e938`.
 It separates completed local runs from runner-ready public benchmarks and planned external evaluations.
 
 Planned rows are not claimed wins. They are the public proof path WaveMind must complete before stronger production claims.
@@ -20,7 +20,7 @@ Planned rows are not claimed wins. They are the public proof path WaveMind must
 | Long-term memory evidence | long-term-agent-memory | implemented | WaveMind: evidence recall@k 1.00, precision@1 1.00, stale suppression 1.00, context saved 0.87, avg latency 6.10, p95 latency 8.99<br>Static vector: evidence recall@k 1.00, precision@1 0.57, stale suppression 0.00, context saved 0.88, avg latency 0.65, p95 latency 0.94 | Use public LoCoMo and LongMemEval for external claims; keep this synthetic scenario for fast regression testing. |
 | BEIR-style open retrieval runner | retrieval | implemented | WaveMind: nDCG@k 0.35, Recall@k 0.48, MRR@k 0.32, precision@1 0.24, avg latency 117.0, p95 latency 256.6<br>Chroma: nDCG@k 0.35, Recall@k 0.47, MRR@k 0.32, precision@1 0.24, avg latency 1.79, p95 latency 2.39<br>Qdrant: nDCG@k 0.35, Recall@k 0.48, MRR@k 0.32, precision@1 0.24, avg latency 17.7, p95 latency 23.3 | Add sentence-transformers runs for SciFact, then add NFCorpus as the second BEIR dataset. |
 | [NoMIRACL Russian retrieval](https://huggingface.co/datasets/miracl/nomiracl) | multilingual-retrieval | implemented | WaveMind: nDCG@k 0.43, Recall@k 0.52, MRR@k 0.49, precision@1 0.41, avg latency 10.2, p95 latency 15.5<br>Chroma: nDCG@k 0.44, Recall@k 0.52, MRR@k 0.49, precision@1 0.41, avg latency 2.60, p95 latency 3.44<br>Qdrant: nDCG@k 0.43, Recall@k 0.52, MRR@k 0.49, precision@1 0.41, avg latency 18.9, p95 latency 24.1 | Run sentence-transformers on the same NoMIRACL Russian split and then add the full MIRACL Russian corpus profile when disk/service capacity allows it. |
-| [LoCoMo evidence retrieval runner](https://github.com/snap-research/locomo) | long-term-conversation-memory | implemented | WaveMind: evidence recall@k 0.39, precision@1 0.24, MRR@k 0.31, context saved 0.00, avg latency 3.95, p95 latency 7.44<br>Static vector: evidence recall@k 0.26, precision@1 0.13, MRR@k 0.19, context saved 0.00, avg latency 1.94, p95 latency 3.87<br>Chroma static: evidence recall@k 0.26, precision@1 0.13, MRR@k 0.18, context saved 0.00, avg latency 7.03, p95 latency 9.74<br>Qdrant static: evidence recall@k 0.26, precision@1 0.13, MRR@k 0.19, context saved 0.00, avg latency 147.6, p95 latency 210.2<br>WaveMind sentence: evidence recall@k 0.55, precision@1 0.33, MRR@k 0.43, context saved 0.00, avg latency 3.44, p95 latency 5.56<br>Chroma sentence: evidence recall@k 0.41, precision@1 0.22, MRR@k 0.30, context saved 0.00, avg latency 4.97, p95 latency 6.30<br>Qdrant sentence: evidence recall@k 0.41, precision@1 0.22, MRR@k 0.31, context saved 0.00, avg latency 124.3, p95 latency 149.7<br>WaveMind Core direct: context saved 0.00, avg latency 4.61, p95 latency 7.61<br>WaveMind + Memory OS direct: context saved 0.00, avg latency 5.14, p95 latency 8.50 | The current direct Memory OS run is admitted but slightly below Core on recall; tune feedback-free policies before claiming LoCoMo uplift. |
+| [LoCoMo evidence retrieval runner](https://github.com/snap-research/locomo) | long-term-conversation-memory | implemented | WaveMind: evidence recall@k 0.39, precision@1 0.24, MRR@k 0.31, context saved 0.00, avg latency 3.95, p95 latency 7.44<br>Static vector: evidence recall@k 0.26, precision@1 0.13, MRR@k 0.19, context saved 0.00, avg latency 1.94, p95 latency 3.87<br>Chroma static: evidence recall@k 0.26, precision@1 0.13, MRR@k 0.18, context saved 0.00, avg latency 7.03, p95 latency 9.74<br>Qdrant static: evidence recall@k 0.26, precision@1 0.13, MRR@k 0.19, context saved 0.00, avg latency 147.6, p95 latency 210.2<br>WaveMind sentence: evidence recall@k 0.55, precision@1 0.33, MRR@k 0.43, context saved 0.00, avg latency 3.44, p95 latency 5.56<br>Chroma sentence: evidence recall@k 0.41, precision@1 0.22, MRR@k 0.30, context saved 0.00, avg latency 4.97, p95 latency 6.30<br>Qdrant sentence: evidence recall@k 0.41, precision@1 0.22, MRR@k 0.31, context saved 0.00, avg latency 124.3, p95 latency 149.7<br>WaveMind Core direct: context saved 0.00, avg latency 4.61, p95 latency 7.61<br>WaveMind + Memory OS direct: context saved 0.00, avg latency 5.14, p95 latency 8.50<br>WaveMind current public protocol: evidence recall@k 0.55, precision@1 0.33, MRR@k 0.43, avg latency 4.88, p95 latency 7.67, ingest avg ms 0.61<br>WaveMind + Memory OS current public protocol: evidence recall@k 0.55, precision@1 0.33, MRR@k 0.43, avg latency 5.99, p95 latency 8.67, ingest avg ms 0.49<br>Mem0 OSS: evidence recall@k 0.50, precision@1 0.26, MRR@k 0.37, avg latency 270.2, p95 latency 293.1, ingest avg ms 217.9<br>Hindsight OSS: evidence recall@k 0.32, precision@1 0.05, MRR@k 0.15, avg latency 320.6, p95 latency 463.1, ingest avg ms 62.3 | Repeat the real-system protocol on LongMemEval-S and add pinned answer-quality readers without presenting retrieval evidence as final answer quality. |
 | [LongMemEval evidence retrieval](https://github.com/xiaowu0162/LongMemEval) | long-term-agent-memory | implemented | WaveMind: evidence recall@k 0.78, precision@1 0.70, MRR@k 0.76, context saved 0.87, avg latency 7.27, p95 latency 9.14<br>Chroma static: evidence recall@k 0.52, precision@1 0.36, MRR@k 0.46, context saved 0.89, avg latency 16.0, p95 latency 18.7<br>Static vector: evidence recall@k 0.52, precision@1 0.36, MRR@k 0.46, context saved 0.89, avg latency 0.08, p95 latency 0.10<br>Qdrant static: evidence recall@k 0.52, precision@1 0.36, MRR@k 0.46, context saved 0.89, avg latency 398.5, p95 latency 432.9<br>WaveMind Core direct: context saved 0.99, avg latency 8.26, p95 latency 12.6<br>WaveMind + Memory OS direct: context saved 0.99, avg latency 8.89, p95 latency 12.4 | The direct run now matches Core retrieval quality; improve maintenance amortization before enabling more aggressive adaptive policies on this dataset. |
 | [LongMemEval evidence 50-query smoke](https://github.com/xiaowu0162/LongMemEval) | long-term-agent-memory | implemented | WaveMind: evidence recall@k 0.92, precision@1 0.76, MRR@k 0.83, context saved 0.87, avg latency 15.3, p95 latency 42.9<br>Chroma static: evidence recall@k 0.60, precision@1 0.26, MRR@k 0.39, context saved 0.89, avg latency 13.3, p95 latency 26.0<br>Static vector: evidence recall@k 0.60, precision@1 0.26, MRR@k 0.39, context saved 0.89, avg latency 0.34, p95 latency 2.14<br>Qdrant static: evidence recall@k 0.60, precision@1 0.26, MRR@k 0.39, context saved 0.89, avg latency 180.2, p95 latency 296.7 | Speed up full LongMemEval reruns by reusing per-question candidate indexes or adding a streaming runner mode. |
 | [ANN index latency curve](https://github.com/erikbern/ann-benchmarks) | index-latency | implemented | WaveMind numpy: Recall@k 1.00, avg latency 1.99, p95 latency 1.45, p99 latency 1.63, build ms 395.1<br>WaveMind quantized: Recall@k 0.93, avg latency 16.3, p95 latency 20.1, p99 latency 28.2, build ms 1019.3<br>WaveMind annoy: Recall@k 0.73, avg latency 3.21, p95 latency 4.17, p99 latency 4.41, build ms 2300.3<br>WaveMind faiss: Recall@k 1.00, avg latency 80.5, p95 latency 2.42, p99 latency 3.87, build ms 458.3<br>Qdrant local: Recall@k 1.00, avg latency 33.8, p95 latency 39.2, p99 latency 44.4, build ms 8336.8 | Tune quantized search kernels, add FAISS on Linux/macOS CI, and test Qdrant service-mode curves beyond 50000 vectors. |

diff --git a/benchmarks/CLUSTER_AUTOSCALE.md b/benchmarks/CLUSTER_AUTOSCALE.md
@@ -1,6 +1,6 @@
 # WaveMind Cluster Autoscale Report
 
-Generated: `2026-07-27T10:52:52Z`.
+Generated: `2026-07-27T13:29:25Z`.
 
 Cluster autoscale evidence is extracted from the checked-in scale-readiness artifact. It proves deterministic shard placement, failure-domain availability, autoscale planning, rebalance planning, operator reconciliation, quorum safety, active-active convergence, field-state CRDT behavior, and the 100M capacity envelope on these fixtures. It is not a real 100M vector-query latency benchmark, managed Kubernetes production run, or independent multi-region SLO.
 

diff --git a/benchmarks/COST_EFFICIENCY.md b/benchmarks/COST_EFFICIENCY.md
@@ -1,6 +1,6 @@
 # WaveMind Cost Efficiency Leaderboard
 
-Generated: `2026-07-27T10:53:35Z`.
+Generated: `2026-07-28T22:43:41Z`.
 
 Measured rows come from checked-in load artifacts. Planned rows are capacity and cost contracts only; they do not unlock production latency or recall claims until the matching benchmark result exists.
 
@@ -17,9 +17,9 @@ Measured rows come from checked-in load artifacts. Planned rows are capacity and
 
 | rank | profile | target class | engine | memories | recall | p99 ms | SLO | cost / 1M queries | monthly cost | source |
 |---:|---|---|---|---:|---:|---:|---|---:|---:|---|
-| 1 | sub_100k-wavemind-numpy-streaming-production_streaming_load_smoke_results | sub_100k | WaveMind numpy-streaming | 10,000 | 1 | 0.515 | pass | $0.694 | $182.5 | `benchmarks/production_streaming_load_smoke_results.json` |
-| 2 | sub_100k-qdrant-sharded-service-streaming-production_streaming_load_qdrant_sharded_smoke_results | sub_100k | Qdrant sharded service streaming | 5,000 | 1 | 3.619 | pass | $0.694 | $182.5 | `benchmarks/production_streaming_load_qdrant_sharded_smoke_results.json` |
-| 3 | sub_100k-wavemind-pgvector-streaming-production_streaming_load_pgvector_smoke_results | sub_100k | WaveMind pgvector streaming | 1,000 | 1 | 7.624 | pass | $0.694 | $182.5 | `benchmarks/production_streaming_load_pgvector_smoke_results.json` |
+| 1 | sub_100k-wavemind-numpy-streaming-production_streaming_load_smoke_results | sub_100k | WaveMind numpy-streaming | 10,000 | 1 | 0.459 | pass | $0.694 | $182.5 | `benchmarks/production_streaming_load_smoke_results.json` |
+| 2 | sub_100k-wavemind-pgvector-streaming-production_streaming_load_pgvector_smoke_results | sub_100k | WaveMind pgvector streaming | 1,000 | 1 | 7.624 | pass | $0.694 | $182.5 | `benchmarks/production_streaming_load_pgvector_smoke_results.json` |
+| 3 | sub_100k-qdrant-sharded-service-streaming-production_streaming_load_qdrant_sharded_smoke_results | sub_100k | Qdrant sharded service streaming | 5,000 | 1 | 16.02 | pass | $1.389 | $365 | `benchmarks/production_streaming_load_qdrant_sharded_smoke_results.json` |
 | 4 | sub_100k-qdrant-service-streaming-production_streaming_load_qdrant_smoke_results | sub_100k | Qdrant service streaming | 1,000 | 1 | 17.9 | pass | $1.389 | $365 | `benchmarks/production_streaming_load_qdrant_smoke_results.json` |
 | 5 | 100k-qdrant-service-production_load_qdrant_100k_tuned_results | 100k | Qdrant service | 100,000 | 1 | 21.26 | pass | $1.389 | $365.02 | `benchmarks/production_load_qdrant_100k_tuned_results.json` |
 | 6 | 1m-qdrant-service-streaming-production_streaming_load_qdrant_1m_tuned_results | 1m | Qdrant service streaming | 1,000,000 | 1 | 26.37 | pass | $2.083 | $547.74 | `benchmarks/production_streaming_load_qdrant_1m_tuned_results.json` |

diff --git a/benchmarks/PUBLIC_MEMORY_COMPETITORS.md b/benchmarks/PUBLIC_MEMORY_COMPETITORS.md
@@ -0,0 +1,24 @@
+# Public Memory Competitors
+
+Official LoCoMo: **5,882 memories**, **1,977 evidence queries**, top-5.
+
+| Engine | Recall@5 | Precision@1 | MRR@5 | Query avg | Query p95 | Ingest avg | Ingest scope |
+|---|---:|---:|---:|---:|---:|---:|---|
+| WaveMind | 0.548 | 0.333 | 0.432 | 4.88 ms | 7.67 ms | 0.61 ms | index_and_persistence_from_precomputed_embeddings |
+| WaveMind + Memory OS | 0.548 | 0.332 | 0.431 | 5.99 ms | 8.67 ms | 0.49 ms | index_and_persistence_from_precomputed_embeddings |
+| Chroma static | 0.408 | 0.219 | 0.305 | 4.12 ms | 4.86 ms | 0.75 ms | local_index_write_from_precomputed_embeddings |
+| Qdrant static | 0.409 | 0.219 | 0.305 | 103.27 ms | 111.45 ms | 0.63 ms | local_index_write_from_precomputed_embeddings |
+| Mem0 OSS | 0.500 | 0.263 | 0.369 | 270.24 ms | 293.08 ms | 217.91 ms | end_to_end_native_embedding_and_persistence |
+| Hindsight OSS | 0.316 | 0.052 | 0.148 | 320.61 ms | 463.14 ms | 62.33 ms | end_to_end_native_embedding_and_persistence |
+
+## Verdict
+
+- Quality winner: **WaveMind** at recall@5 `0.548`.
+- Fastest average local query: **Chroma static** at `4.12 ms`.
+
+## Boundaries
+
+- WaveMind, Chroma, and Qdrant ingest starts from shared precomputed embeddings. Mem0 and Hindsight ingest includes their native embedding and persistence work.
+- This is LoCoMo retrieval evidence on one local machine. It is not final answer quality, hosted-service throughput, or an architecture-only comparison because real systems use their pinned native embedding stacks.
+- Source commit: `4b6d7dd3e938dc04663bd7ba2c2356d5e7d1fe62`.
+- Dataset SHA-256: `79fa87e90f04081343b8c8debecb80a9a6842b76a7aa537dc9fdf651ea698ff4`.
diff --git a/benchmarks/agent_impact_results.json b/benchmarks/agent_impact_results.json
@@ -1,7 +1,7 @@
 {
   "schema": "wavemind.agent_impact_leaderboard.v1",
-  "generated_at": "2026-07-28T17:57:37Z",
-  "source_ref": "d011eb848e03",
+  "generated_at": "2026-07-28T22:43:41Z",
+  "source_ref": "4b6d7dd3e938",
   "claim_boundary": "Agent-impact rows come from checked-in benchmark artifacts. They show behavioral lift on the configured tasks; they do not claim general agent success outside the listed scenarios.",
   "summary": {
     "benchmark_count": 8,

diff --git a/benchmarks/benchmark_artifact_audit.json b/benchmarks/benchmark_artifact_audit.json
@@ -1,13 +1,13 @@
 {
   "schema": "wavemind.benchmark_artifact_audit.v1",
   "status": "pass",
-  "checked_at": "2026-07-28T17:57:38Z",
-  "generated_at": "2026-07-28T17:57:37Z",
-  "source_ref": "d011eb848e03",
+  "checked_at": "2026-07-28T22:43:43Z",
+  "generated_at": "2026-07-28T22:43:41Z",
+  "source_ref": "4b6d7dd3e938",
   "workflow_run_id": null,
   "refresh_profile": "local",
   "max_age_days": 8.0,
-  "age_days": 1.5496840277777777e-05,
+  "age_days": 2.698244212962963e-05,
   "implemented_count": 45,
   "runner_ready_count": 3,
   "planned_count": 5,