v0.36.0

0.36.0 (2026-04-18)

Features

• abstraction: archive escape hatch for chronically-demoted abstractions (18ec2c5)
• abstraction: archive escape hatch for chronically-demoted abstractions (45b61a9)
• abstraction: gate synthesis on substrate-change fingerprint (2b92ead)
• abstraction: gate synthesis on substrate-change fingerprint (209020b)
• abstraction: streak-based archive supersedes age-based decay exit (917edb3)
• abstraction: streak-based archive supersedes age-based decay exit (083d0c1)
• add Gemma 4 E2B spoke inference server (6a3c409)
• consolidation: resurrect archived patterns on re-emergence (24c9943)
• consolidation: resurrect archived patterns on re-emergence (0c27890)
• continuous learning Phase B — curriculum generation (#391) (e5a5d59)
• continuous learning Phase C — training trigger & orchestration (#391) (5bb0094)
• continuous learning Phases B+C — curriculum generation & training trigger (#391) (28860cf)
• dreaming: INFO logs for four silent-drop gates (865a71d)
• dreaming: INFO logs for four silent-drop gates (c9f40bd)
• dreaming: rotate replay sample via recent-ID ring buffer (1a97250)
• dreaming: rotate replay sample via recent-ID ring buffer (b8ae0a7)
• enable continuous learning by default (#391) (ac2bf1b)
• Gemma 4 E2B spoke training — 25/25 schema compliance (ab6ba40)
• SetSpokeTensorF32 — push F32 weights with auto-quantization to native type (780808f)
• SPLICE API for CRISPR-LM (edit/status/tensor/complete) + Gemma 4 template fix (c77f749)
• SPLICE API for CRISPR-LM + Gemma 4 chat template fix (79289a6)
• training data assembly for continuous learning (#391) (99d3c28)

Bug Fixes

• abstraction: bump MaxTokens 200 -> 500 for principle/axiom synthesis (252d14c)
• abstraction: bump MaxTokens 200→500 for principle/axiom synthesis (ef0ffd4)
• abstraction: concept-overlap gate on findSimilarAbstraction (604bf33)
• abstraction: concept-overlap gate on findSimilarAbstraction (cbabbbd)
• abstraction: lower MinStrength default 0.7 → 0.5 to unblock principle synthesis (d89fa4a)
• abstraction: lower MinStrength default 0.7 → 0.5 to unblock principle synthesis (98a8408)
• add GBNF grammar for episode synthesis (588f802)
• add GBNF grammar for episode synthesis to fix "Untitled session" failures (151296c)
• address all PR #397 review items from Caleb (e52ee2a)
• agents: axiom grounding, pattern dedup, forum transition reporting (1b7ba03)
• agents: axiom grounding, pattern dedup, forum transition reporting (0cff8b5)
• consolidation: bound salience with a 1.0 ceiling in decay (d80e297)
• consolidation: bound salience with a 1.0 ceiling in decay (daff51e)
• consolidation: break pattern-dedup super-attractor via concept-overlap gate (c1a3a39)
• consolidation: break pattern-dedup super-attractor via concept-overlap gate (d50c860)
• consolidation: concept-overlap gate on second-stage dedup (3372dc9)
• consolidation: concept-overlap gate on second-stage dedup (f9135fd)
• consolidation: sample large clusters + raise MaxTokens for identifyPattern (cfae847)
• consolidation: sample large clusters + raise MaxTokens for identifyPattern (cbb5434)
• correct bool array type in Gemma GGUF export + update EXP-30 verdict (45515b2)
• curriculum_runs stored time.Time.String() output, unparseable on read (1266e46)
• curriculum_runs stored time.Time.String() output, unparseable on read (f1eec70)
• Gemma 4 spoke training produces garbage due to use_cache=False (b1eaa8e)
• llm: drop misleading "busy" prefix on embedded provider cancel (48b3b19)
• llm: drop misleading "busy" prefix on embedded provider cancel (450c886)
• move advisory board personas from local memory to .claude/skills/ (2dd06b0), closes #398
• move advisory board personas into repo (5c4e5bb)
• prevent MCP processes from loading GPU models into VRAM (8bd86a8)
• prevent MCP processes from loading GPU models into VRAM (a575622)
• store: age-bound training breaker so stale failures expire (0198153)
• store: age-bound training circuit breaker so stale failures expire (6e76dc2)
• Task Scheduler is logon-only, manual start uses PID-file daemon (07a605f)
• training crash loop safety + episode synthesis quality (f0a8a95)
• training crash loop safety + episode synthesis quality (#391) (75dc713)
• type-filtered recall surfaces recent memories first (86673ea)
• type-filtered recall surfaces recent memories first (#394) (5af04de)
• use gemma-4-E2B-it (not base) in stress test, add EXP-31 results (845c9cb)
• Windows daemon survives reboots, add dashboard update button (b73424f)
• Windows daemon survives reboots, add dashboard update button (356c570)

Performance Improvements

• BetaQ all-RQ4 101 tok/s, embedded Gemma 4 deployment (578313a)

What's Changed

• fix: type-filtered recall surfaces recent memories first by @CalebisGross in #395
• fix: move advisory board personas into repo by @CalebisGross in #399
• feat: Gemma 4 E2B spoke training — 25/25 schema compliance by @CalebisGross in #400
• refactor: lean advisory board skill by @CalebisGross in #401
• docs: add llama.cpp inference benchmarks to EXP-31 by @CalebisGross in #402
• fix: add GBNF grammar for episode synthesis by @CalebisGross in #405
• feat: continuous learning Phases B+C — curriculum generation & training trigger (#391) by @CalebisGross in #406
• fix: training crash loop safety + episode synthesis quality by @CalebisGross in #407
• fix: prevent MCP processes from loading GPU models into VRAM by @CalebisGross in #408
• feat: SPLICE API for CRISPR-LM (edit/status/tensor/complete) + Gemma 4 template fix by @CalebisGross in #409
• docs: refresh CLAUDE.md + rules for Gemma 4 production & crispr-lm coupling by @CalebisGross in #410
• fix: curriculum_runs stored time.Time.String() output, unparseable on read by @CalebisGross in #411
• fix(consolidation): break pattern-dedup super-attractor via concept-overlap gate by @CalebisGross in #412
• fix(consolidation): sample large clusters + raise MaxTokens for identifyPattern by @CalebisGross in #413
• fix(consolidation): concept-overlap gate on second-stage dedup by @CalebisGross in #414
• fix(abstraction): concept-overlap gate on findSimilarAbstraction by @CalebisGross in #415
• fix(abstraction): lower MinStrength default 0.7 → 0.5 to unblock principle synthesis by @CalebisGross in #416
• feat(dreaming): INFO logs for four silent-drop gates by @CalebisGross in #417
• fix(llm): drop misleading "busy" prefix on embedded provider cancel by @CalebisGross in #418
• fix: Windows daemon survives reboots, add dashboard update button by @jkbennitt in #397
• fix(abstraction): bump MaxTokens 200→500 for principle/axiom synthesis by @CalebisGross in #419
• feat(abstraction): archive escape hatch for chronically-demoted abstractions by @CalebisGross in #420
• feat(abstraction): streak-based archive supersedes age-based decay exit by @CalebisGross in #421
• fix(agents): axiom grounding, pattern dedup, forum transition reporting by @CalebisGross in #422
• feat(consolidation): resurrect archived patterns on re-emergence by @CalebisGross in #425
• fix(store): age-bound training circuit breaker so stale failures expire by @CalebisGross in #426
• feat(dreaming): rotate replay sample via recent-ID ring buffer by @CalebisGross in #427
• feat(abstraction): gate synthesis on substrate-change fingerprint by @CalebisGross in #428
• fix(consolidation): bound salience with a 1.0 ceiling in decay by @CalebisGross in #429
• chore(main): release 0.36.0 by @appsprout-release-please[bot] in #396

Full Changelog: v0.35.0...v0.36.0