Zaxy 2.1.0: Cognitive Defaults, Measured First

Zaxy 2.1.0 ships the additive 2.1–2.3 agent-experience and cognitive-memory feature set and promotes two defaults that earned the flip through their measurement lanes.

What Is New

The 2.1–2.3 surface is entirely additive on the unchanged immutable Eventloom log: MCP tool listing profiles and umbrella tools, checkout token budgets with a cache-stable consolidated prefix, the replayable salience reinforcement ledger, encoding-specificity cues, the experimental memory_feeling_of_knowing metamemory pre-check, procedure mining, vector re-embedding, and the internal agent-experience and cognitive measurement lanes. No Eventloom envelope changes, no projection migrations, no removed or renamed tools. The full upgrade path, including both default flips and their opt-outs, is documented in docs/migration.md.

Default Flips (and Their Evidence)

MCP_TOOL_PROFILE now defaults to core. The internal tool-adoption lane measured the listed tool surface at 8,165 estimated tokens under full versus 1,344 under core — an 83.5% reduction — with the memory_checkout front door listed first. Profiles change listing only: dispatch is never filtered, so every tool stays callable by name, and memory_capabilities reports the available-but-unlisted set. Opt out with MCP_TOOL_PROFILE=full or zaxy serve --profile full.

RETRIEVAL_PROFILE now defaults to cognitive. The internal forgetting lane measured exact cold-start parity (zero-reinforcement corpora rank byte-identically to plain retrieval), no-recall-loss 1.0 (attenuated memories always reachable by explicit query, and labeled), pin/authority exemptions 1.0, and ranking lift 1.0 versus 0.0 for plain. The cognitive profile composes the same local retrieval stack as local_fast plus salience-blended ranking, cue blending, and the personalized-PageRank graph walk. The graph-walk lane separately measured multi-hop bridge lift on 6/6 entity-bridge queries (versus 0/6 under plain ranking) with zero single-hop regressions (4/4 direct queries retained at rank 1 in both arms). Opt out with RETRIEVAL_PROFILE=local_fast. Settings that leave the profile unset but customize embedding/reranker/scoring knobs keep resolving to the custom profile exactly as before.

Who is affected: clients that snapshot tool listings should refresh against the core listing or pin MCP_TOOL_PROFILE=full; users with reinforcement history (or graph structure the walk can traverse) may see ranking shifts, while cold-start users see identical ranking.

What Did Not Flip

• ENCODING_GATE_ENABLED stays false: the write-time gate's effect is unmeasured, and unmeasured defaults do not flip.
• The feeling-of-knowing surface stays experimental: its calibration lane Brier margin over the base-rate predictor is +0.001–0.006 — real but too thin to promote.
• Vector quantization stays none pending a production-dimension re-run of the vector-scale lane, although int8 is the strongest promotion candidate measured (recall@10 0.9938, 0.14x memory, 1.8x faster than exact at 10^5).
• VECTOR_ANN_THRESHOLD is raised from 50000 to 1000000, keeping the Kuzu HNSW path effectively opt-in: the vector-scale lane measured it below exact search on both recall (0.90 versus the 0.95 bar at 10^5) and latency, with non-reproducible index builds. The path remains available by lowering the threshold explicitly; the default will come down only after per-query projected-graph reuse and recall tuning land with lane evidence.

Fixes

A feeling-of-knowing verdict boundary was corrected: a 3-term query with one known term scores exactly the possible threshold (0.6 * (1/3) = 0.2), but binary floating point evaluated it just below and mislabeled it unlikely. Threshold comparisons are now epsilon-tolerant; scores are unchanged.

Upgrading

pip install --upgrade zaxy-memory
zaxy doctor

To keep 2.0 behavior exactly: MCP_TOOL_PROFILE=full and RETRIEVAL_PROFILE=local_fast. Everything else upgrades without action. See docs/migration.md and docs/configuration.md.