Skip to content

Monomind v2.5 — The Second Brain Grows Up

Choose a tag to compare

View all tags
@nokhodian nokhodian released this 18 Jul 19:47
v2.5.1
de1b648

Monomind v2.5 — The Second Brain Grows Up

Covering releases 2.3.0 through 2.5.1 · Now published: v2.5 (2.5.1)

When we shipped 2.2, the Second Brain was a promising sketch: a local memory store your sessions could query. Over the last release wave — 2.3.0 through 2.5 — it has grown into what it was always meant to be: local-first knowledge that is cross-project, live, and now available to entire agent organizations. Along the way we rebuilt the memory engine from the ground up, put the new modules through a 49-agent adversarial review, and gave the Org Runtime the operational tooling it needed to run unattended.

Here's the story in three arcs.

The Second Brain, everywhere

2.3.2 laid the foundations: heading-aware chunking (with § section prefixes so you can see exactly which part of a document a hit came from), session-start reindexing, and provable pruning via cleanup --data.

2.3.3 made it ambient. A warm embedding model in the dashboard server answers /api/knowledge/search in roughly 60ms, and relevant [SECOND_BRAIN] excerpts are injected into substantive prompts automatically — each hit labeled with (semantic) or (keyword) provenance so you know why it surfaced. Telemetry records timings only, never prompt text.

2.4.0 broke the project boundary. Your personal knowledge now lives in one global brain at ~/.monomind/global-brain, with zero-decision routing: ingests from outside a project route there automatically (or force it with --global), and retrieval everywhere merges both stores with project-first ranking and clearly labeled global hits.

2.5.0 closed the loop: org agents can now read the Second Brain. A new knowledge_search tool gives every agent in an org access to project and global knowledge, with every lookup audited on the org bus. Documents are ingested live — a saved file is searchable about five seconds later, courtesy of the warm model, with a silent fallback to session-start reindexing on platforms without recursive file watching. The dashboard gained a scope selector and global badges to match.

Orgs that remember, report, and ask

2.3.0 was a capability wave for Org Runtime v2, aimed at one thing: running orgs you can actually operate.

  • Observability: org logs --follow for live tailing; org report for tokens vs. budget, produced assets, and crashes.
  • Memory across runs: runs record an org_complete outcome, and each new run gets a cross-run briefing from history.jsonl — orgs stop rediscovering what the last run already learned. 2.3.2 extended this with org_recall, letting agents query prior-run memory directly.
  • Humans in the loop, headlessly: agents can raise questions that you answer later via org questions / org answer — no need to babysit a terminal.
  • Resilience and ergonomics: crashed sessions restart automatically with backoff, org run --dry-run previews a run, and org create --template plus org validate make new orgs cheap to scaffold and safe to trust.

A leaner, hardened engine

2.3.1 replaced the memory engine outright. LanceDB — and roughly 600MB of native dependencies — is gone, replaced by local SQLite (better-sqlite3, with a sql.js WASM fallback) and local MiniLM embeddings. The rewrite also fixed real bugs: empty semantic search results, keyword search that only matched exact phrases, and namespace leaks. A paraphrase golden-set evaluation now runs in CI to keep retrieval quality honest.

2.3.4 was the hardening pass. We ran a 49-agent adversarial review against the swarm and memory stack and fixed all 33 findings it produced — including a critical cleanup --data rule that would have deleted live memory stores, and a window in which org messages could be silently lost during session restarts. Alongside: code-fence and CRLF fixes in the chunker, a UNIQUE(namespace, key) constraint, TTL exclusion from retrieval, and a small quality-of-life guarantee — every failed CLI command now prints its reason.

2.5.1 rounds out the wave as a patch on 2.5.0.

Upgrading

npm i -g monomind@latest
# or run without installing:
npx monomind@latest

Full details

Every change above is documented in the CHANGELOG. Issues, questions, and contributions: github.com/monoes/monomind.

Monomind coordinates, Claude Code creates — and now, it remembers.

Assets 2
Loading