Skip to main content
    Norditco logo
    All engineering posts
    Published November 14, 2026Norditco Engineering

    Norditco Brain, Institutional Memory Without Privileged Wiki Pages

    • product
    • architecture
    • agents
    • memory

    Norditco runs an autonomous product studio. Product Owner, Scout, Builder, Growth, Security, and adversarial review agents already share one hybrid ops board. Agents still paid a tax on every session. They rediscovered architecture facts, repeated the same safe formatting steps, and either under-loaded context or pasted too much.

    We did not solve that by making a shared wiki the new system of record for authority. We built Norditco Brain: a shared memory and capability-evolution service that existing agents call at major workflow boundaries. The first version proved the gateway and a dedicated machine-memory space. Brain 2.0 hardened retrieval, evaluation, promotion, private networking, and the human Overview so the studio can use memory in production without inventing a second control plane.

    What Brain is (and is not)

    Brain is not a new manager agent. Human-approved policy and the studio controller stay in charge. Brain provides:

    1. Recall of compact, verified context for a task
    2. Remember / reflect for durable lessons with evidence
    3. Pattern detection when the same low-risk work repeats
    4. Sandboxed promotion of Tier-0 helpers into Git
    5. Eval gates before promotion when thresholds are required
    6. Tool Scout that rejects unverifiable tools by default
    7. One human Overview so growth is reviewable in minutes
    8. Offline reconcile when the laptop cannot reach the VPS
    Architecture diagram. Expand for a larger view with zoom.

    Separation that a shared wiki still respects

    On many shared-wiki deployments, every logged-in site member can edit every space. A dedicated Brain space is therefore a structural boundary, not a confidentiality boundary by itself.

    We keep four kinds of state apart:

    KindHome
    Policy baselineGit (versioned JSON policy)
    Durable knowledgeDedicated machine-memory wiki space
    Runtime index / auditLocal SQLite under the studio data directory
    Executable helpersGit capability scripts

    Rules that matter in practice:

    • Never execute code copied from wiki pages
    • Never store secrets in Brain pages
    • Brain writes go through an allowlisted adapter that refuses human documentation spaces
    • Retrieved text is data. Prompt-injection strings are quarantined, not obeyed
    • Git security and QA gates win when memory disagrees
    • Embeddings stay on the control plane host. Remote embed APIs are blocked by cost and data-handling policy

    Human product documentation stays in the human docs space. Agents get machine-oriented records in the Brain space. Humans get a single Brain Overview with attention items, alerts, a Mermaid (and optional Graphviz) tree, and capability status.

    Recall at boundaries, not every tool call

    Loading the entire Brain into every prompt would raise cost and dilute signal. Recall runs at mandatory boundaries such as Product Owner intake, portfolio decisions, Builder start/resume, Growth planning, opportunity evaluation, QA/release checks, and end-of-run retrospectives.

    Each skill’s early step is a compact CLI call:

    npm run cli -- recall --boundary builder_start_resume --role builder \
      --intent 'How should Lab apps be deployed?' --project lab --compact
    

    Each recall returns:

    • Trusted policy snippets from Git
    • A small set of scored memories (typical pilot estimates around a few hundred tokens of summaries, not a space dump)
    • Freshness / conflict warnings
    • Stable IDs so agents can fetch a full record only when needed
    • A retrieval_mode so operators can see hybrid versus keyword degradation

    Expired market signals are excluded with an explicit warning. Near-duplicate summaries collapse to one memory with both evidence trails retained in the audit journal.

    Hybrid retrieval that can degrade honestly

    Brain 2.0 indexes memories with FTS5 for keywords and sqlite-vec for local vectors. Query-time fusion uses reciprocal rank fusion (RRF). Embeddings come from a local model served only on the control-plane host. The daemon uses a short keep-alive so agent sessions do not thrash unload and reload.

    When the local embedder is down, recall falls back to keyword FTS5 and reports a degraded retrieval mode. That is intentional. Agents still get Git policy and lexical matches; they do not silently call a remote embedder.

    Self-evolution with a short leash

    When agents repeat the same safe sequence, Brain fingerprints the work. A candidate appears after roughly three similar successes in two weeks, or about ten minutes of avoidable toil per week.

    Promotion is gated by risk tier and by evals:

    • Tier 0: pure local helpers, no network, no secrets, no production writes. Sandbox test, then activate within standing authorization when evals pass.
    • Tier ≥1: file writes, tracker or wiki side effects, deploys, new privileges. Human approval; agents cannot grant themselves rights.
    • require_evals: the eval suite (retrieval, security, memory, restart, agent, automation) must pass; critical security graders stay at 100%.

    The first pilot helper was intentionally boring: formatting a memory JSON record into wiki storage HTML. It lives in Git, refuses credential-shaped payloads, and was activated only after static analysis and a sandbox run.

    Tool Scout reviews candidate tools with a checklist. Unknown provenance and incompatible licenses are rejected by default. Local candidates stay awaiting human approval. Agents cannot self-authorize a new dependency.

    Human Overview without a second product UI

    There is no Brain SaaS console. Operators run:

    npm run cli -- dashboard
    

    Artifacts land in the local studio data directory (overview.md, attention.json, Mermaid/DOT). The wiki Brain Overview page is refreshed from those files via the allowlisted adapter. Attention rows name severity, evidence, recommended action, and consequence if ignored. Failures, disputed memories, pending Tool Scout decisions, and offline jobs surface first. Low-value noise is capped.

    Private mesh without publishing Brain

    Laptop and VPS share an encrypted control-plane mesh. The tunnel is split so only mesh traffic uses it. Brain, the local embedder, and product apps stay on the host; they are not published on the mesh interface.

    Production mesh bring-up runs through an allowlisted control-plane installer after configuration is staged. Keys live under a local config directory and never enter Git.

    When the mesh or SSH path is down, Brain queues allowlisted offline jobs (heartbeat, connectivity probe, wiki outbox flush) and drains them idempotently when connectivity returns. Arbitrary shell from a payload is refused.

    Controlled pilot evidence

    A Builder boundary recall for Lab deploy returned the verified static-product lesson first, with Git policy snippets, at a measured token estimate far below a full wiki dump. A mesh networking lesson was remembered, formatted with the Tier-0 helper, synced to the machine-memory space, and ranked first for a mesh-intent recall. Offline connectivity probes drained over the live mesh. We do not invent percentage savings against “paste the whole space.” The benefit is structural: boundary recall, allowlisted sync, and gated promotion.

    Operations and recovery

    Brain supports an emergency stop flag that forces policy-only recall, append-only change events, outbox-style wiki sync keys, and DB backup/export. Restoring a backup file is enough to rehydrate the local index after a bad run; wiki pages remain versioned and must not be blindly overwritten when a human edited them first.

    Session start is deliberately boring: start the local embedder if hybrid recall is needed, confirm connectivity, load the current work summary from the hosted board, then call Brain at the workflow boundary. That habit matters more than another dashboard.

    Why this shape

    We already had discovery SQLite, wiki and work-board adapters, diagram publishing, and a roster of skills. Brain reuses those pieces instead of inventing a second control plane. The success metric is not page count. It is useful recall, fewer repeated mistakes, safer delivery, and a human who can see a week of growth on one Overview page.

    Related: Autonomous product discovery, Hybrid ops and the AI inbox.