This release includes 4 breaking changes for platform teams planning a safe upgrade.
✓ No known CVEs patched in this version
Topics
+14 more
Summary
AI summaryKnowledge graph removed and API functions now require noun_tags.
Full changelog
TagCascade Retrieval — Benchmark-Validated Architecture
Replaces knowledge graph with TagCascade retrieval validated on LoCoMo benchmark.
Highlights
- 85.8% on LoCoMo (non-adversarial, end-to-end answer accuracy)
- +23 points over raw ingestion (76.6% vs 53.0%, p<0.0001)
- TagCascade retrieval: tags-first cascade + CE reranking, raw CE score (no Wilson blend)
- Two-lane retrieval: 4 summaries + 4 facts = 8 memories per context injection
- Atomic fact extraction: LLM extracts facts via MCP tools
- CE via ONNX: multilingual mmarco-mMiniLMv2, CPU-only, no PyTorch
- Poison resilient: -2.6 to -4.2 points with 1,135 adversarial memories
Breaking Changes
- Knowledge graph removed (replaced by TagCascade)
score_memoriesandrecord_responsenow requirenoun_tagsparameteradd_to_memory_banknow requiresnoun_tagsparameter- Haiku removed from sidecar auto-fallback (must explicitly configure)
New Commands
roampal sidecar test— validate scoring model response format
Full Details
See release notes for complete changelog.
Paper: "Beyond Ingestion: What Conversational Memory Learning Reveals on a Corrected LoCoMo Benchmark" (Logan Teague, April 2026)
Benchmark: roampal-labs
Breaking Changes
- Knowledge graph removed (replaced by TagCascade)
- `score_memories` and `record_response` now require `noun_tags` parameter
- `add_to_memory_bank` now requires `noun_tags` parameter
- Haiku removed from sidecar auto-fallback (must explicitly configure)
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About roampal-ai/roampal-core
Outcome-based persistent memory for AI coding tools. Memories that help get promoted, memories that mislead get demoted. Works with Claude Code and OpenCode via hooks + MCP.
Related context
Beta — feedback welcome: [email protected]