The matter that remembers itself.
capturing everything a complex matter accumulates (chronology, evidence, procedural rules, known traps) into a governed system the whole team can query and trust.
Matter knowledge normally lives in one or two heads. It degrades under fatigue, disappears on handover, and produces inconsistency between documents drafted weeks apart. In a multi-year matter, memory is the single biggest unmanaged risk.
A long-running matter is a knowledge problem wearing a legal costume: a contested chronology spanning years; exhibits under active renumbering; multiple witnesses, each with a mapped posture; causes of action with different elements and standards; a set of known drafting traps; and hard procedural rules about what each document type may and may not do.
Every new working document has to be consistent with all of it. And every generic AI tool, handed the file, fails in the same predictable ways: inventing citations, importing foreign law into an Australian matter, paraphrasing where only verbatim quotes are safe, and confidently repeating errors from earlier drafts.
Decompose · Build · Validate · Supervise
The matter is factored into its load-bearing structures: the spine chronology, with dates verified against source and contested dates marked as contested; witness posture maps; the exhibit register; the themes register; and the case’s organising propositions, against which every draft can be tested.
Each structure becomes grounded context for a domain-specific drafting assistant, wrapped in guardrails that encode professional discipline as system behaviour rather than user vigilance:
- Verbatim or flagged. No party’s words quoted unless present in the source; paraphrase always labelled as paraphrase.
- Jurisdiction-locked. Australian authority and citation style; foreign doctrine structurally blocked, not merely discouraged.
- Known traps pre-armed. Contested references and previously identified errors trigger automatic flags whenever any draft touches them.
- Role-bounded. Every output carries its status on its face (a working draft for the legal team to settle) and never predicts outcomes or makes calls reserved to the practitioner.
The system is tested against the matter itself: known answers, known traps, known contested points. It goes into use only when it reliably flags the traps instead of falling into them.
The system maintains an explicit register of its own limits (documents not yet provided, facts not yet verified) and says so when asked to work near them, instead of filling gaps with plausible fiction. Every material development updates the system, so it evolves with the matter rather than snapshotting it.
| Before | After |
|---|---|
| Matter knowledge in one or two heads | Explicit, queryable, consistent for the whole team |
| Generic AI: fabrication and foreign-law risk | Grounded system: source-anchored, jurisdiction-locked |
| Known traps re-triggered draft after draft | Traps encoded once, flagged automatically forever |
| Every document starts from a blank page | Every document starts from the verified state of the matter |
| Handover = weeks of re-reading | Handover = access |
This is the use case the other four sit inside. Courts have sanctioned practitioners for filing AI-fabricated citations; the difference between that outcome and this one is not the model. It’s the engineering. A governed case system is what “using AI in legal practice” looks like when it’s built to survive scrutiny: the firm’s own knowledge as the specification, guardrails as system behaviour, a human owning every call, and an audit trail that satisfies the practitioner’s own conduct obligations.
Every long matter in your firm is already a knowledge system. The only question is whether it’s engineered, or just remembered.