The Myco roadmap.

11 brain_* tools over MCP, backed by your own Postgres 16 + pgvector. Boots with one docker compose up, zero configuration, seeded workspace included.

BM25 search and document ingestion work with zero API keys, and since v1.1.0, semantic search is keyless too. Add an Anthropic key only for the most accurate graph extraction.

Identical content is rejected on write, so re-ingesting a folder never multiplies your memory.

Every accepted fact links back to its source, ask brain_why and get the evidence chain, with how it was extracted and how many sources support it.

Point mycobrain-ingest at a local folder or a GitHub repo and your text files become searchable, source-traceable memory (binaries and noise directories are skipped).

Entity extraction, entity resolution (duplicate names collapse into one node), and entity-to-entity relationships. Build it fully locally with Ollama, no API key, or with Anthropic for the most accurate graph. Direction-aware extraction since v1.1.0 (npm run test:direction).

Shipped in v1.1.0: semantic search via local embeddings (nomic-embed-text on Ollama), alongside the existing OpenAI path. Keyless full-text and keyless semantic search both work with zero hosted dependencies (npm run test:local-embeddings).

The extraction worker proposes new entity kinds and relationship types it observes in your data, surfaced by brain_stats, and, under explicit opt-in (BRAIN_SCHEMA_AUTO_PROMOTE=1), proposals corroborated across enough independent documents auto-promote into the live catalogs with a full audit trail. Promotion stays manual by default, and strict curation mode always wins (npm run test:schema-promotion).

A fact's confidence rises as independent sources corroborate it and falls when contradicted. On single-valued predicates the contradicted fact is superseded, closed and weakened, never silently overwritten, with the supersession recorded in the claims ledger. brain_why shows source counts, confidence trend, and superseded history (npm run test:compounding).

The LongMemEval harness ships in-repo (evals/longmemeval): 73.6% end-to-end QA accuracy on the complete 500-question oracle subset with 100% evidence-retrieval recall, reader gpt-4o-mini, judge gpt-4o. Run it yourself.

Point mycobrain-ingest at an OpenAI or claude.ai data export (--from chatgpt-export / --from claude-export) and every conversation becomes provenance-tracked, deduplicated, searchable memory, the active branch only, re-imports safe (npm run test:export-import).

The extraction prompt is directed (subject → predicate → object) with worked examples, lifting llama3.2:3b to 86% directed accuracy (12/14) and ~80% edge survival (11–12/14) on the gold fixture, measurable with npm run test:direction.

mycobrain-doctor (one-command preflight), mycobrain-review (audited human-in-the-loop curation), mycobrain-rest (read-only search + why over HTTP), and import-first mycobrain-onboard, the day-two ergonomics of running your own memory.

Agent identity comes only from the API key, secrets are redacted from the audit table, and workspace RLS binds under the least-privilege brain_app role (the agency kit ships it; the default local superuser is single-workspace use).

A documented, reproducible local model setup for higher-quality graphs, so the keyless path gets closer to the hosted one.

Today an agency’s shared playbook is read-only by convention. Next: enforce read-only at the Postgres role level, so the shared layer can’t be mutated even by mistake.

Underway: retrieval is now measured on the full longmemeval_s distractor haystack (recall@5 89.2% keyless, 91.6% with the recency reranker), not just the evidence-only oracle subset. Next: more categories and harder conditions.

First-class connectors beyond files, repos, and chat exports.

Hosted, multi-tenant Myco for teams that would rather not operate Postgres themselves. Currently waitlist-only, not generally available.