Changelog

0.25.0-alpha.onnx.2 (2026-07-08)

Experimental feature build (second in the memory-embeddings series). Everything here is gated behind Memory.Embeddings.Enabled, off by default — without opting in, behavior is identical to the mainline beta. Not published to the beta channel; install only by exact pin: NETCLAW_VERSION=0.25.0-alpha.onnx.2.

Memory (Experimental)

• Hybrid semantic recall with an absolute relevance floor — automatic pre-turn recall now unions FTS5 lexical and embedding-cosine candidates (identical policy gates for both), fuses scores with recency decay, and enforces a gold-set-calibrated minimum-similarity floor: when nothing relevant exists, nothing is injected. Zero-injection turns are normal and healthy.
• Cross-encoder relevance gate — a 22 MB int8 reranker (ms-marco-MiniLM-L-6-v2, hash-pinned) scores each floor survivor against the query and drops weak matches; measured out-of-sample at 86.8% zero-injection accuracy with 98.3% recall retention. Follows Memory.Embeddings.Enabled; degraded mode falls back to floor-only recall, never blocks a turn.
• Model-documented query prefix + manifest-carried calibration — recall queries now embed in arctic-embed's documented retrieval mode; each allowlisted model pins its prefix and calibrated floor together, and Memory.Recall.MinCosineSimilarity follows the active model's calibration unless explicitly overridden. Measured on the production gold set: F0.5 +73%, recall@3 2.8×, zero-injection accuracy 2.1× vs the unprefixed configuration.
• int8 arctic embedder is the new default model — Snowflake's pre-quantized model_uint8.onnx (105 MB vs 416 MB fp32, ~1.7× faster, measurably better retrieval quality with the prefix). fp32 and mxbai remain allowlisted as explicit choices.

Upgrading from 0.25.0-alpha.onnx.1 with embeddings enabled

• The default model id changes to snowflake-arctic-embed-m-int8. On first daemon start the warmup gap-repair sweep re-embeds your corpus under the new model automatically (recall degrades to lexical-only for unembedded documents until coverage completes); netclaw memory backfill-embeddings --force does it in one pass. Existing fp32 vectors are left in place and untouched; netclaw doctor will note the mixed-model rows until you re-backfill. Original memory content is never modified.