Skip to content

Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors

Rust

docs: publish 3.3 retrieval benchmark

AkitaOnRails251b4d80b9a1
diff --git a/docs/benchmarks/README.md b/docs/benchmarks/README.md
index 0efc68e8..f6d35303 100644
--- a/docs/benchmarks/README.md
+++ b/docs/benchmarks/README.md
@@ -13,6 +13,8 @@ cargo run --release -p ai-memory-eval -- retrieval --fetch
| date | dataset | mode | overall hit@5 | file |
|---|---|---|---|---|
+| **2026-10-29** | LongMemEval-S (v1) | **zero-llm FTS + observation stream (3.3.0 release)** | **0.951** | [longmemeval-s-3.3.0-release.md](longmemeval-s-3.3.0-release.md) |
+| **2026-10-29** | LongMemEval-S (v1) | **local embeddings + observation stream (3.3.0 release)** | **0.945** | [longmemeval-s-3.3.0-release.md](longmemeval-s-3.3.0-release.md) |
| **2026-10-26** | LongMemEval-S (v1) | **zero-llm FTS + observation stream (3.2.6 release)** | **0.951** | [longmemeval-s-3.2.6-release.md](longmemeval-s-3.2.6-release.md) |
| **2026-10-26** | LongMemEval-S (v1) | **local embeddings + observation stream (3.2.6 release)** | **0.945** | [longmemeval-s-3.2.6-release.md](longmemeval-s-3.2.6-release.md) |
| **2026-10-25** | LongMemEval-S (v1) | **zero-llm FTS + observation stream (3.2.4 release)** | **0.951** | [longmemeval-s-3.2.4-release.md](longmemeval-s-3.2.4-release.md) |
@@ -37,6 +39,13 @@ cargo run --release -p ai-memory-eval -- retrieval --fetch
| 2026-09-01 | LongMemEval-S (v1) | zero-llm, stopword-filtered FTS (prior) | 0.668 | [longmemeval-s-2026-09-01-fts.md](longmemeval-s-2026-09-01-fts.md) |
| 2026-09-01 | LongMemEval-S (v1) | zero-llm (pre-2.0 FTS) | 0.617 | [longmemeval-s-2026-09-01.md](longmemeval-s-2026-09-01.md) |
+The 3.3.0 minor-release run found FTS identical to 3.2.6 at every accuracy
+and recall cutoff and kept local `hit@5` exact. Local `hit@10` and
+`recall@10` each moved by 0.007, slightly outside the documented overall
+noise band, and are tracked in #1854. See the
+[3.3.0 release report](longmemeval-s-3.3.0-release.md) for provenance, every
+slice, and the explicit threshold assessment.
+
The 3.2.6 patch-release run found FTS identical to 3.2.4 at every accuracy
and recall cutoff. Local embeddings kept the same `hit@5` and `hit@10`
overall, while `recall@10` improved. See the
diff --git a/docs/benchmarks/longmemeval-s-3.3.0-release.md b/docs/benchmarks/longmemeval-s-3.3.0-release.md
new file mode 100644
index 00000000..e02a5318
--- /dev/null
+++ b/docs/benchmarks/longmemeval-s-3.3.0-release.md
@@ -0,0 +1,75 @@
+# LongMemEval-S release benchmark: 3.3.0
+
+The 3.3.0 release benchmark runs all 470 answerable LongMemEval-S questions
+through the shipped hook capture pipeline. It measures the default zero-LLM
+FTS path and the in-process local embedder, both with the production-default
+observation stream enabled.
+
+## Provenance
+
+- code: both modes used `c92047ed3b5f639aab615ae4e33f3310c52c0843`,
+ the frozen 3.3.0 release candidate before this report.
+- dataset: `longmemeval_s`, sha256
+ `08d8dad4be43ee2049a22ff5674eb86725d0ce5ff434cde2627e5e8e7e117894`.
+- hardware: AMD Ryzen 9 7950X3D (32 threads). The two modes ran serially;
+ no build or test workload overlapped the local-embedding run.
+- commands, after building `ai-memory` and `ai-memory-eval` in release mode:
+
+ ```bash
+ ai-memory-eval retrieval \
+ --server-bin target/release/ai-memory \
+ --concurrency 6 --http-timeout-secs 900 \
+ --server-env AI_MEMORY_RETRIEVAL__OBSERVATION_STREAM=true \
+ --server-env AI_MEMORY_EVAL_SKIP_WIKI_CHECKPOINTS=1
+
+ ai-memory-eval retrieval \
+ --server-bin target/release/ai-memory \
+ --concurrency 6 --http-timeout-secs 900 \
+ --embeddings local \
+ --server-env AI_MEMORY_RETRIEVAL__OBSERVATION_STREAM=true \
+ --server-env AI_MEMORY_EVAL_SKIP_WIKI_CHECKPOINTS=1
+ ```
+
+- run dirs: `evals/runs/2026-10-29T21-20-20-570928323Z-retrieval` (FTS)
+ and `evals/runs/2026-10-29T21-21-17-842507936Z-retrieval` (local).
+
+## Results
+
+FTS matches 3.2.6 exactly at every accuracy and recall cutoff. Local
+embeddings keep the primary `hit@5` metric exact. Local `hit@10` and
+`recall@10` each moved by 0.007, roughly three questions and just outside the
+documented +/-0.005 overall noise band. The result is published rather than
+re-run until favorable; #1854 tracks the per-question investigation.
+
+| mode | hit@1 | hit@3 | hit@5 | hit@10 | recall@5 | recall@10 | p50 | p95 | mean context tokens |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| FTS 3.2.6 | 0.857 | 0.930 | 0.951 | 0.972 | 0.893 | 0.937 | 11 ms | 73 ms | 534.8 |
+| **FTS 3.3.0** | **0.857** | **0.930** | **0.951** | **0.972** | **0.893** | **0.937** | **11 ms** | **73 ms** | **534.8** |
+| local 3.2.6 | 0.766 | 0.904 | 0.945 | 0.977 | 0.879 | 0.947 | 157 ms | 363 ms | 535.2 |
+| **local 3.3.0** | **0.762** | **0.909** | **0.945** | **0.970** | **0.876** | **0.940** | **141 ms** | **330 ms** | **534.5** |
+
+Local `hit@5` moved on two small slices while remaining exact overall:
+
+| slice | n | 3.2.6 | 3.3.0 | delta |
+|---|---:|---:|---:|---:|
+| knowledge-update | 72 | 1.000 | 1.000 | 0.000 |
+| multi-session | 121 | 0.975 | 0.975 | 0.000 |
+| single-session-assistant | 56 | 0.893 | 0.893 | 0.000 |
+| single-session-preference | 30 | 0.867 | 0.900 | +0.033 |
+| single-session-user | 64 | 0.969 | 0.953 | -0.016 |
+| temporal-reasoning | 127 | 0.913 | 0.913 | 0.000 |
+
+## Latency and ingest
+
+FTS query latency stayed at 11/73 ms p50/p95. Local query latency moved from
+157/363 ms to 141/330 ms. FTS ingested 278,946 items in 55.2 seconds. Local
+embeddings ingested the same items in 1,624.1 seconds. Ingest wall time is
+load-sensitive and is not a release gate. Both runs returned zero
+raw-observation fallback hits.
+
+## Verdict
+
+The primary retrieval gate passes: FTS is exact and local `hit@5` is exact.
+The top-ten movement is small but outside the pre-registered overall noise
+band, so it is published here and tracked in #1854 instead of hidden or used
+to tune this release.