Skip to content

Changelog

Product updates and release notes for Myco Brain.

v1.3.02026-10-15

The memory layer, open source, in one command

  • One command: npm install -g @mycobrain/mcp-server, then mycobrain-setup memory. It sets up local Postgres with pgvector, or takes a Supabase, Neon, RDS or other Postgres connection string, plus a starter template: agency, design studio, architecture or landscape firm, solo consultant, or custom.
  • Claude Code and Codex get hooks that recall at session start and on each prompt and capture finished sessions. Cursor, Windsurf and Claude Desktop get MCP with two memory tools, brain_memory_context and brain_memory_write.
  • Facts keep their source and supersede instead of overwrite, with history kept (mycobrain-memory history). Houses keep clients apart, and mycobrain-memory binding add scopes an agent to one client.
  • A background worker harvests finished sessions with your own Claude or Codex subscription and embeds locally with Ollama, no API keys. Dream proposes merges, dossiers and patterns; it is off by default and nothing changes until you approve it. The document engine still ships, now optional.
v1.2.92026-06-24

Guided install, memory provenance by tool, and a self-check that talks

  • mycobrain-setup: a guided, self-healing install. A nine-step, consent-driven walkthrough replaces the cryptic fresh-machine failure. It checks prerequisites with a fix for each, wires every client it detects, verifies the connection for real (semantic search present and a live write test), offers a one-tap import of your ChatGPT or Claude history, and runs the doctor. Experts skip the asking with --yes, --all, or --client.
  • Memory provenance by tool: each connected app gets its own identity, so brain_recall_memory and brain_context_pack show which tool a memory came from, and a cross-tool recall is credited (recalled from Cursor's memory). brain_stats and brain_self_check show a per-tool breakdown so compounding across tools is visible.
  • Surfacing modes and the self-check that talks: brain_set_mode sets how loudly Myco surfaces (silent, ambient, or audit, and silent by default). brain_self_check is a pull-only call that reports what is working, what is awaiting your approval, and any problem with a concrete fix.
  • Additive only: two new tools and additive fields on existing ones, with no breaking changes to any brain_ tool and no migrations.
v1.2.62026-06-17

Adjudication-engine agent contract, live doctor, contract eval

  • Agent contract rewritten to embody the product thesis: the program writes the facts, not the model. It teaches the source-first write ladder (ingest a source, then propose a claim, then a private save_memory) and the engine mechanics: compounding confidence across independent sources, supersede-don't-overwrite on conflict, provenance via brain_why, and propose to review to promote.
  • brain_save_memory is honestly fenced as a private scratchpad, never workspace truth. The contract is single-sourced across the runtime, the pasteable CLAUDE.md / .cursorrules / AGENTS.md manual, the tool descriptions, and the website, with a CI drift guard.
  • mycobrain-doctor now verifies live, not just config: it pings Ollama, confirms the model is pulled, and runs a real embed and generation, with a --fix that pulls missing models. brain_ingest returns an honest extraction receipt (graph vs search-only).
  • Agent-contract eval harness (npm run eval:contract): 20 adversarial scenarios scoring tool-routing adherence, so the contract is measured and regression-guarded. No tool-contract changes, no migrations.
v1.2.52026-06-17

One-command install, repo onboarding, value surfacing

  • npx @mycobrain/install: one command detects your MCP client and writes the config for Claude Code, Claude Desktop, Cursor, Codex, or Windsurf, then offers to onboard. Connecting your agent went from hand-editing JSON per client to a single command.
  • Onboarding now indexes your own repo (opt-in): on a fresh brain it asks before indexing the current project, then a new session recalls your own code with the file cited.
  • You can see it remember: brain_recall_memory and brain_context_pack surface a decaying 'recalled from your memory' credit line, plus a once-per-session stats greeting and save milestones.
  • mycobrain-ingest --watch-downloads: opt-in watcher that auto-imports a ChatGPT or Claude export zip the moment it lands in ~/Downloads, identifying which by the files inside the zip. No tool-contract changes, no migrations.
v1.2.42026-06-16

Reliability and security hardening

  • Extraction-worker durability: a lease-based reaper re-queues chunks stranded after a crash, so documents no longer stick in pending_llm with zero entities (#70).
  • Contradiction and supersession hardening: concurrent contradictions are serialized, conflict matching is predicate-normalized, and the claims ledger is de-duplicated (#72). seen_count and brain_why now count distinct source documents (#73).
  • Workspace-scoped dynamic type catalogs so one workspace's vocabulary can't leak into another (#71, migration 052), plus a resolved transitive form-data security advisory (#77). No tool-contract changes.
v1.2.32026-06-16

Benchmark accuracy: recall@5 confirmed on the full 500-question run

  • Documentation-accuracy release, no code changes: the recency-reranker recall@5 first measured on a 100-question sample was re-run on the complete 500-question LongMemEval longmemeval_s set before launch.
  • Full-500 figures: 89.2% keyless default (hybrid, 446/500) → 91.6% with the reranker (temporal, 458/500). The default scored higher at full scale than the sample showed.
  • Reproduced two independent ways with zero drift (retrieval is deterministic). Every number we publish has to reproduce on the full set. When a sample and a full run disagree, the full run wins and we say so.
v1.2.22026-06-16

Keyless recency reranker, better retrieval, no API key

  • Keyless recency reranker: brain_search(reranker: 'recency') reorders by final = 0.7·relevance + 0.3·recency_norm, deterministic, no API key, no network call. 15/15 reranker tests.
  • Recall@5 on the full 500-question LongMemEval longmemeval_s set: 89.2% keyless default (hybrid) → 91.6% with the reranker (temporal). Verified two ways, zero drift; reproduce with the eval harness.
  • Eval robustness: embedder inputs sanitized so one bad batch can't silently skew results.
  • README accuracy: edge-survival figure corrected to 0% → ~80% (11–12/14) (86% is the direction metric).
v1.2.12026-06-13

Import your ChatGPT / Claude history, agent instructions, security & RLS hardening

  • Import your ChatGPT / Claude history: mycobrain-ingest --from chatgpt-export (or --from claude-export) turns a data export into provenance-tracked, deduplicated, searchable memory, active branch only, re-imports safe, brain_why traces every fact (npm run test:export-import).
  • Out-of-box agent instructions: the MCP server ships a usage contract to every client at init, so agents know when to recall, save, and cite without per-project setup.
  • Operator tooling: mycobrain-onboard (import-first, sandboxed tour, dry-run --reset-demo), mycobrain-doctor (one-command preflight), mycobrain-review (audited human-in-the-loop curation), mycobrain-rest (read-only search + why over HTTP).
  • Security: agent identity comes only from the key, caller-supplied workspace/agent args are ignored for brain_* keys; secrets redacted in the audit table; docker ports bind to 127.0.0.1 (npm run test:sharing).
  • RLS least-privilege hardening under the brain_app role (migration 20260615000050); local-model extraction hardening; brain_save_memory works out of the box; direction figure re-measured to 86%.
v1.2.02026-06-12

Compounding confidence, full dynamic schema, per-object privacy, in-repo benchmark

  • Compounding confidence: a fact's confidence rises with independent corroboration, falls on contradiction, contradicted facts are superseded via the claims ledger, never silently overwritten (npm run test:compounding).
  • Full dynamic schema: corroborated proposals auto-promote into live catalogs under explicit opt-in, 3 independent docs at ≥0.8 confidence, audit trail, strict mode always wins (npm run test:schema-promotion).
  • Per-object privacy: private documents readable only by their creating agent across all six read tools (npm run test:sharing).
  • LongMemEval harness in-repo: 73.6% end-to-end QA accuracy on the complete 500-question oracle subset, 100% evidence-retrieval recall, reader gpt-4o-mini, judge gpt-4o. Run it yourself: evals/longmemeval.
  • With this release, everything on the original launch roadmap has shipped.
v1.1.02026-06-11

Keyless semantic search, direction-aware graphs, dynamic schema (phase 1), strict curation

  • Local (keyless) vector search via Ollama nomic-embed-text, semantic search now works with no hosted API key at all (npm run test:local-embeddings).
  • Direction-aware relationship extraction: subject → predicate → object with worked examples, lifts llama3.2:3b direction accuracy from ~79% to ~86% on the gold fixture (npm run test:direction).
  • Dynamic schema phase 1: the extraction worker proposes new entity kinds and relation types from your data, surfaced by brain_stats, promotion stays manual (npm run test:dynamic-schema).
  • Strict curation mode: BRAIN_REQUIRE_HUMAN_REVIEW=1 routes every extraction through the review queue, the LLM proposes, the human decides (npm run test:strict-mode).
  • Seeded relationship-type catalog: eight canonical predicates ship as the dynamic-schema baseline.
  • No breaking changes to any brain_* tool contract.
v1.0.02026-06

First public open-source release

  • First public release of Myco Brain, open source under Apache-2.0, at github.com/thegoodguysla/myco-brain and on npm as @mycobrain/mcp-server.
  • Self-hosted MCP server exposing 11 brain_* tools: context packs, search, provenance, graph traversal, ingestion, agent memory, and workspace stats.
  • Postgres 16 + pgvector schema (42 migrations) with a seeded default workspace, docker compose up works with zero configuration.
  • Keyless full-text (BM25) search and document ingestion, no API keys needed to get started.
  • Content-hash deduplication on ingest, plus queryable provenance via brain_why with evidence summaries.
  • Bulk ingest CLI (mycobrain-ingest) for local folders and GitHub repos.
  • Knowledge graph: entity extraction, entity resolution, and entity-to-entity relationships, fully local via Ollama, or Anthropic for the most accurate graph.
  • Reproducible dedup + provenance benchmark in examples/benchmark/.
  • CI that boots the real Docker stack and runs all tools end-to-end.
  • Managed cloud remains waitlist-only and is not part of this release, self-hosting is the default.
v0.1.02026-05-20

Public launch prep, sales page, infrastructure

  • Sales page rebuilt with pain-led hero, 9-offer structure, and Railway CTA.
  • Marketing landing page served at mycobrain.dev root via hostname detection.
  • Separate marketing and app surfaces via hostname routing.
  • Onboarding email sequence automated for first 7 days.
  • Retention tracking infrastructure rebuilt with migrations, API routes, and cron jobs.
  • Beta signup API alias route for waitlist.
  • Graph and agents nav restored with auth fallback.
  • Security headers (X-Frame-Options, CSP) added for app.mycobrain.dev.
v0.0.22026-05-15

Connector stability, entity extraction, and ingestion hardening

  • Entity extraction cron wired to main, activates relationship discovery automatically.
  • Connector bind update SQL typing regression resolved.
  • Ingestion source coverage audit added; async fetch paths stabilized.
  • Advisory synthesis fallback broadened for relations and describer.
  • Subscription tracking migration aligned with production schema.
  • CORS helper added for retention routes.
  • CI gracefully skips external sync when running against localhost.
v0.0.12026-05-01

Core platform: MCP server, knowledge graph, ingestion pipeline

  • MCP server with the brain_* tool suite: context pack, search, ingest, graph traversal, memory, provenance.
  • Postgres 16 + pgvector schema with workspace-scoped access.
  • Deterministic write path, facts deduplicated by content hash and traceable back to their sources.
  • Python workers for ingestion, advisory extraction, connectors, schema evolution.
  • Next.js dashboard and admin apps for knowledge exploration and onboarding.
  • CLI onboarding with cloud and local/offline modes.
  • Budget governor compaction logic for MCP server.
  • SSL self-heal for all database query paths.
RSS Feed
Changelog, Myco Brain