neuralmind.uk / Release Notes — NeuralMind v4.0.0 Edit this page on GitHub ↗

Release Notes — NeuralMind v4.0.0

Release Date: 2026-09-16
License: MIT
Compare: v3.13.0 → v4.0.0


What’s New

Lean-Context-Inspired Memory Modules (v3.13.0 → v4.0.0)

Five new modules built from lean-ctx patterns:

Module Purpose Status
context_budget Fixed token budget with L3→L2→L1→L0 progressive trimming ✅ Shipped
session_summaries Periodic semantically-recallable digests of agent sessions ✅ Shipped
graph_traversal Learned co-access edges that strengthen with use ✅ Shipped
read_dedup Content-hash dedup + LRU cache eliminates redundant reads ✅ Shipped
cognition_loop Background knowledge consolidation: reinforce, decay, promote, prune ✅ Shipped

Medical Retriever Integration

NeuralMind.query() now routes prose/book projects to a specialized BM25 + terminology-expansion retriever. Enables peptide-book, clinical-manuscript, and long-form prose indexing alongside code.

Compliance Hardening

Site + Marketing


Benchmarks (Reproducible)

Metric Value Command
Gold-file recall 93.75% mean / 79–100% per repo python -m evals.public.run
Token reduction vs pasting 45–257× Same
Requests backend 45.7× fewer tokens Same
Click backend 107.7× fewer tokens Same

Benchmark is reproducible:

pip install neuralmind && python -m evals.public.run

Every miss is published. No cherry-picking.


Upgrade Notes

No breaking changes. Existing NeuralMind users:

pip install --upgrade neuralmind
neuralmind wakeup .   # picks up new modules automatically

Memory layer (decision storage + commit-level invalidation) is not in this release — it’s the next build target per the Memory Layer spec.


Assets


Full Changelog: CHANGELOG.md