Monomind v2.5.3 — Memory upgrade, monodesign goes standalone, npm packaging fix

Monomind v2.5.3

This release covers everything since v2.5.1 (v2.5.2 shipped to npm but was never tagged or released on GitHub, so this note includes it). Headline items: a packaging fix affecting every npm-installed user, a memory upgrade inspired by cognee, and monodesign graduating into its own standalone package.

Packaging fix (affects every npm install)

Three internal packages (hooks, mcp, routing) were still declared under the old @monomind npm scope, which was never published, while their siblings (memory, monobrowse, monodesign) had already moved to @monoes. They were listed as optional dependencies with a wildcard version, so npm install completed without error but skipped all three every time. The practical effect for anyone who installed via npm: the background-worker system (security scans, git tracking, KG consolidation, health/performance monitoring, and 8 more workers), the standalone MCP server, and semantic task routing were all absent. Monorepo developers were unaffected since a local workspace link covered the gap.

Fixed: @monoes/hooks, @monoes/mcp, and @monoes/routing are now published at 1.0.0 under the correct scope (two of them were also missing their compiled output in the tarball, which would have kept them broken even after the scope rename). Verified after publishing: a fresh install resolves the real registry packages, and monomind hooks worker list lists all 14 workers from the built CLI binary.

Memory upgrade: closed feedback loop and conversation knowledge graph

A five-phase port of ideas from cognee (the open-source AI memory platform), adapted to monomind's existing TypeScript stack, with no new dependencies and no new database.

• Closed feedback loop: memory entries carry a feedback weight (updated from usage ratings via exponential moving average) and a frequency weight (incremented on every real use). Both feed into future search ranking, but only for genuine semantic matches; keyword-fallback results are excluded, so a highly-rated stale entry cannot outrank a relevant one it has no real similarity to. Rating application is idempotent per task.
• Conversation and org knowledge graph: entities, relationships, and distilled rules extracted from agent sessions and org runs, with deterministic name-based identity so the same entity merges across sessions instead of duplicating. Org coordinators get a new org_learn tool, with an entity glossary injected into their briefing to reduce near-duplicate entities. Interactive sessions get memory_kg_ingest, memory_kg_search, and memory_kg_rollback. Every write is provenance-tagged so a single run's contribution can be rolled back.
• org_remember: agents can write memory scoped to the whole org or private to their own role; org_recall searches both.
• Rule-based query router with reciprocal-rank fusion: a lightweight regex classifier picks the retrieval surface (document chunks, knowledge graph, rules, or memory) per query, and results from multiple surfaces are merged by rank.
• KG consolidation: the existing consolidate background worker now merges stale entity descriptions using neighborhood facts already in the graph, mechanically and fully locally.
• Observability: a doctor check for knowledge-graph size, an org memory CLI (stats, search, rules, rollback), and a new panel on the orgs dashboard showing entities, relations, rules, and a glossary.
• Several MCP tool descriptions referenced internal controllers that did not exist; those are corrected, and the causal-edge tool now writes a real, traversable graph edge instead of an inert record.
• Verified on a real two-run org: org_learn populated the graph, the next run's briefing included the resulting glossary, org_recall returned the learned relationship, and the feedback weight updated exactly as designed after a successful outcome.

monodesign is now a standalone package

Monodesign (frontend design intelligence: tokens, antipattern detection, the monodesign skill) has moved out of this repo into its own package and project, published independently as @monoes/monodesign.

• Live mode proven end to end via monobrowse: boot, wrap, variant generation, accept or carbonize, discard, manual-edit routing, and steer sessions.
• New design fix subcommand: deterministic codemods for 8 mechanically safe antipattern rules, with a detect-fix-re-detect verification loop.
• Design memory: critique snapshots now flow into monomind's memory system and the Second Brain automatically; a new recall surface backs the polish flow.
• 5 new detector rules, taking the total from 46 to 51: missing focus-visible states, small touch targets, hover-only affordances, images without dimensions, and dark-scheme contrast blind spots, wired across all four detection engines.
• Native monobrowse driver for browser-based detection; puppeteer is now the fallback path instead of primary.
• Per-file attribution comments removed; legal attribution consolidated into the license and a notice file. Monodesign is maintained as an independent project going forward.

monobrowse fixes

• Connect sessions survive across CLI invocations; closing a session no longer leaves a headless Chrome process running.
• Scans past an occupied default port instead of failing outright.

Org runtime and dashboard

• Per-role model hints in org templates, for scheduled-org cost tuning.
• Org chart re-flows into a serpentine grid with a full-page expand view; long edges curve, and bubbles get safe padding.
• Dashboard motion switched to compositor-friendly properties throughout, removing layout-triggering animations and bounce easing.
• Fixes: org migrate no longer leaves a stray literal text artifact in the reports-to field; org mark-complete sends the dashboard auth token; org list checks process liveness; the dashboard token now propagates and self-registers correctly across known projects on startup.
• Fixed a completion and deadlock edge case in the org runtime, and added a per-session latch on the graph gate.

Mastermind skills

• The flat skills layout has been retired in favor of one top-level skill directory per skill, resolving cases where a command referenced a sub-skill name that did not resolve.
• techport is now a directory-based skill.

Also in this release

• An environment variable now disables the monograph pre-bash gate for headless and CI runs.
• Document ingestion no longer skips the monodesign directory, so its critique snapshots are searchable in the Second Brain.
• Fixed two CI test flakes that only timed out under full-suite parallel load; per-test timeout raised for those specific cases rather than changing the global default.

Full diff: v2.5.1...v2.5.3

Install with: npx monomind@latest doctor --fix
Or: npm install -g @monoes/monomindcli@latest