Changelog

0.27.0-beta.1 (2026-08-27)

This beta opens the semantic memory cycle. NetClaw now embeds and recalls memories locally on your machine with ONNX — no cloud embeddings, no external services. Recall gets a relevance gate that keeps weak matches out, and netclaw doctor watches your embedding model's health.

Local semantic memory (ONNX)

A new Netclaw.Embeddings assembly gives every cross-session memory a local embedding, searched and gated on-device.

• Embed at write time, recall by meaning. Memories are embedded when stored, nominated by kNN during curation, and found through hybrid vector + lexical recall (#1749).
• A cross-encoder keeps weak matches out. OnnxCrossEncoderScorer applies a post-floor relevance gate so vague hits don't slip into recall (#1749).
• Local, pinned models. Defaults are snowflake-arctic-embed-m-int8 for embedding and ms-marco-minilm-l-6-v2 for the gate. Both come from a pinned allowlist — NetClaw never loads an arbitrary model (#1749).
• Faster and lighter at rest. The int8 embedder used about 57% less steady-state RSS than fp32 and ran about 1.7 times faster (#1749).
• Provisioning you can control. Models download on first daemon start, alert through the operational channel on failure, and can be pre-provisioned for a fully offline start (#1749).
• New CLI and health tooling. netclaw memory backfill-embeddings embeds an existing corpus, and netclaw doctor reports embedding model health (#1749).

Small fixes

• Fixed release-note parsing so version metadata resolves cleanly (#2067).

Upgrade note: embeddings default to enabled. Expect roughly 800 MB total RSS on the reference configuration, and plan for model access on first start. Operators can disable embeddings or pre-provision the models for offline use.