Skip to main content

​
Overview

By default the Genesis Team Wiki is dormant — agents have no durable, cross-thread memory of what they learn. Once enabled, agents capture the facts and conventions they pick up in a conversation and apply them automatically in later, unrelated threads, with no one re-teaching them. Two capabilities come online:
  • On-demand capture — an agent (or a Flow) deliberately saves a learning about a table, view, or dataset.
  • Background extraction (the “Librarian”) — a background process reads completed human chat threads, distills the durable facts and conventions, and attaches them to the relevant data assets automatically.
Once a learning is captured, it surfaces in any future thread: when an agent looks up a table in the Context Graph, the attached conventions come back with it and shape the answer.
Example. A user teaches an agent, over a long thread, that revenue must come from a specific summary view (not a raw activity table), that one report type must be excluded to avoid double-counting, and that a blank country code means a specific internal category. Days later, in a brand-new thread, the agent computes that revenue correctly — applying all three rules — without being told again.

​
Correcting a stale learning

For the supported workflow to refine a Team Wiki learning or retire it from the default final-filtered overlays, use its current identifier from wiki_read, set the appropriate category with wiki_edit, and follow Context Tools → Team Wiki tools. If a retry follows an earlier same-text edit that did not persist its category or privacy class, Genesis reapplies the missing state instead of silently deduplicating the retry. To remove specific values from a learning and restore its metadata privacy class, provide that class explicitly; see the Context Tools reference for the edit behavior and stale-replay rules.

​
How it works

There are three moving parts:
  1. Capture — on-demand (wiki_write) or automatic (the Librarian extracts learnings from completed chat threads on a schedule).
  2. Anchoring — every learning is attached to a qualified name, typically a data asset (a table or view) in the Context Graph. If the anchor is not catalogued in the graph, the learning remains available through a synthesized, learnings-only Team Wiki page and browseable qualified-name branch.
  3. Retrieval — when an agent searches the data context, attached learnings come back automatically. This retrieval path is always on and does not depend on any agent-facing wiki tool.
For SQLite assets, retrieval also checks the historical connection-less anchor used by older learnings and past SQL records when the asset is requested through its current graph identity. Existing records therefore remain available without being migrated; the tool reference documents the exact alias behavior.
Enabling extraction is necessary but not sufficient for graph-backed retrieval: the data your agents talk about should be catalogued in the Context Graph so learnings can attach to the scanned asset and participate in its normal context. If an anchor is not yet in the graph, active authored learnings remain readable from the Team Wiki, but the page is marked as ungrounded until the asset is scanned.

​
Prerequisites

Confirm all four before enabling:

​
Supported LLM providers for the extractor

The extractor uses your deployment’s active LLM provider — the one selected under AI Providers. The prefix in wiki.extractor_model names the provider, and it must be the active one: Pick the model from the dropdown in Advanced Settings — it lists the active provider’s models, already qualified with the right prefix. A value naming a provider other than the active one is refused when you save it.
Switching the active provider pauses the wiki until you re-pick its models. A model chosen under the old provider still names that provider, so the wiki stops calling it rather than keep spending on a provider you moved away from. Genesis warns you when you make the switch. Choose new values for wiki.extractor_model (and wiki.correction_classifier_model, if set) from the new provider to resume background extraction.

​
H2 correction classifier model

When wiki.correction_classifier_enabled is true, optionally set wiki.correction_classifier_model (default "") to choose the model used by the H2 correction scan. The value is read live when the scan runs, so no restart is required. It must be a non-Claude model; leave it empty to use WIKI_CORRECTION_CLASSIFIER_MODEL. With neither set, the scan stops with an error that names both settings. See Advanced Settings → Team Wiki / Librarian settings for the full configuration reference. The H2 correction scan watches both ways an agent can rely on a Team Wiki statement: a search result that surfaces it and a statement-tagged wiki_read or hierarchical browse event. After the reliance, the scan checks the next user turn in the same thread for a contradiction. If the same statement is surfaced and read before that turn, Genesis classifies the reliance once rather than spending twice or counting the same correction twice. Reads without a thread are not scanned.

​
Enable it

All settings live in Config → Advanced Settings under the wiki section. See Advanced Settings → Team Wiki / Librarian settings for the full reference of every flag; the steps below are the minimal path to turn on durable learnings.
1

Set the extractor model

Set wiki.extractor_model to a model from your active LLM provider (see above). Do this first — enabling extraction without a model causes errors every cycle.
2

Enable the Librarian

Set wiki.librarian_enabled to true. This arms background extraction (takes effect on the next cycle — no restart).
3

Enable the master gate

Set wiki.write_enabled to true. This is the master gate for the whole wiki feature.
4

Restart the server once

Restart so agents see the wiki tools. Background extraction and automatic learning-retrieval already work without it, but the restart completes the experience. Batch all your changes first, then restart once.
5

(Optional) keep settings safe on shared deployments

Leave wiki.admin_trace_bodies_enabled set to false on any shared/multi-tenant deployment. Optionally enable wiki.admin_tools_enabled (also needs a restart) to get operator inspection tools for auditing what was captured.

​
Do I need to restart?

Even before a restart, once the gates are on the Librarian begins extracting, and the learnings it produces surface to agents automatically through Context Graph search. The restart is needed only so agents can use the direct wiki tools (read a full wiki page, write on demand) — not for capture or for the automatic recall that delivers most of the value.
A restart re-checks two environment kill-switches — WIKI_LIBRARIAN_DISABLED and GENESIS_RUN_WIKI_LIBRARIAN. If either is set to disable the Librarian in your service spec, the background loop will not start regardless of wiki.librarian_enabled. Your Advanced-Settings values persist across the restart.

​
Verify it works

1

Teach it

In a substantial conversation (several turns), have an agent work with a real, catalogued table and state the conventions explicitly — the right source, any exclusion rules, what special values mean.
2

Wait at least an hour

The Librarian deliberately ignores threads that are still active; a thread must be idle for about 60 minutes before it is eligible for extraction, so it never extracts mid-conversation.
3

Open a fresh thread

Ask the same question plainly, with no rules given. The agent should apply the conventions and produce the right answer.
4

Confirm capture directly

Ask the agent to look up that table in the data context — it should now report attached learnings (a non-zero learnings count) quoting the conventions. With the admin inspection tools enabled, you can audit captured learnings directly.

​
What to expect

  • ~1-hour lag for the durable layer. Because of the idle-settling rule, a freshly taught convention will not appear in the wiki for about an hour. In the meantime, an agent can still recall a prior thread directly if asked (“bring up that report we did”) — that path is immediate and needs none of this setup.
  • Conversations must be substantial. Very short threads (a couple of messages) are skipped as too thin to extract from.
  • Ungrounded anchors remain browsable. Active learnings about qualified names not in the Context Graph are served as synthesized Team Wiki pages and branches, marked as ungrounded until the corresponding asset is scanned.
  • Extraction uses the LLM. Each eligible thread costs an extraction call; the Librarian backs off when there is nothing new.

​
Troubleshooting

The thread may not be idle 60+ minutes yet, may be too short, or the table may not be catalogued in the Context Graph. Wait past the settling window, use a substantial thread, and scan the connection into the graph.
wiki.librarian_enabled is on but wiki.extractor_model is empty. Set the extractor model.
A WIKI_LIBRARIAN_DISABLED or GENESIS_RUN_WIKI_LIBRARIAN environment kill-switch is set in the service spec. Clear it and restart.
wiki.write_enabled was flipped but the server has not been restarted yet. Restart once.
An environment override may be outranking the database value. Check the service spec for a matching env override and remove it.

Advanced Settings

Full reference for every wiki.* setting, the model picker, and wiki.db storage and recovery.

Digital Twin and Context Graph

Browse Team Wiki alongside the graph, and confirm your data is catalogued so learnings can anchor.