Memory class architectures for firm-scoped legal intelligence

Almost every legal AI product now describes its system as "learning from your firm." In observed practice that claim resolves to one of three storage patterns: conversation history retained in the model context window, embeddings of past chat sessions cached for retrieval, or user inputs pooled (sometimes silently) into shared base-model fine-tunes. None of these patterns are memory architectures. They are storage choices framed as intelligence.

The cognitive-science distinction between episodic and semantic memory (Tulving, 1972; Squire, 1992) and the procedural-vs-declarative split (Anderson, 1996) maps onto a problem legal AI has not solved: how does a system retain the firm's accumulated judgment (which redlines a partner refused, which jurisdictional defaults a client requires, which procedural sequences resolved a class of matters) in a form that is governable, scoped, and revocable. Raw chat history cannot. Embeddings of chat sessions cannot. Silent fine-tunes can, but only in a form that fails on attorney-supervision tests (ABA Model Rule 5.3; ABA Formal Op. 512, 2024).

FirmIQ™ is the FlowCounsel™ layer that distills attorney review signal into firm-scoped patterns the system retrieves on the next matter. It is a memory architecture, not a context cache. The remainder of this paper specifies its three classes, the capture discipline, the attribution mechanism, the promotion state machine, and the retrieval boundary. The framing extends an earlier argument in the FlowCounsel blog post "Why Legal AI Memory Is a Systems Problem, Not a Prompt Problem" (FlowCounsel, April 2026).

The three-class taxonomy (episodic, semantic, procedural) is grounded in cognitive-science research on long-term memory systems (Tulving, 1972; Squire, 1992; Anderson, 1996) and has informed several recent agent-systems architectures (Park et al., 2023; Packer et al., 2023; Wang et al., 2023). The translation for legal workflow requires each class to answer a different question, carry a different lifecycle, and run under a different governance discipline.

The strongest learning signal in legal work is the final diff between what the specialist produced and what the attorney approved. That diff carries the firm's judgment in a form the system can attribute. Character-by-character keystroke telemetry is rarely the high-signal artifact and is almost always the wrong privacy posture (FlowCounsel, "Legal AI Has a Surveillance Problem Before It Has a Regulation Problem," May 2026).

FirmIQ™ default capture is structured: artifact version before review, artifact version after review, structured diff, changed span locators, reviewer comments, approval state, reviewer identity and role, matter and client metadata, specialist contract version, context manifest id, and verifier results available at the time. The diff-from-feedback pattern parallels the verbal-reinforcement learning used in Reflexion (Shinn et al., 2023), with the verbal feedback replaced by structured attorney review.

Editing session duration, review sequence, comment resolution order, and clause movement are captured selectively when a specific workflow benefits, but raw keystroke surveillance is excluded by default. The signal-to-burden ratio is poor and the privacy posture conflicts with the confidentiality obligations under ABA Model Rule 1.6.

A trajectory intelligence extractor reads the workflow run after approval and identifies the earliest meaningful divergence between what the specialist produced and what the attorney approved. That divergence is the candidate learning event. The pattern is analogous to Reflexion's episodic reflection on task feedback (Shinn et al., 2023), narrowed to structured artifact diffs rather than free-text reflection.

Worked example: a specialist drafted a liability cap at 2x fees. The attorney changed the data-breach liability to uncapped. The attribution is not "attorney edited the draft." The attribution is "the draft failed to account for sensitive PII triggering the firm's uncapped-liability exception." The resulting candidate becomes a firm rule scoped to the firm, the client, the contract type, and the data/privacy context, with evidence captured: artifact diff, reviewer identity, approval state, context manifest id.

The extractor does not mutate active FirmIQ™ rules. It produces candidates. Promotion requires the explicit governance described in §5.

Promotion from candidate to active firm pattern requires explicit governance. This constraint maps directly to ABA Model Rule 5.3, which obligates lawyers to ensure nonlawyer assistance (including software tools, per ABA Formal Op. 512, 2024) operates under supervision compatible with the lawyer's professional obligations. A learning system that silently promotes one attorney's edit into a firm-wide rule moves outside that supervisory boundary.

Allowed promotion sources: approved artifacts, repeated approved edits across multiple matters, validated outcome reviews, explicit operator-authored rules, and explicit inversion events. A single edit on one matter is not enough to change firm doctrine.

FirmIQ™ does not load all firm memory into every run. Retrieval is shaped by firm, client, matter, practice area, jurisdiction, artifact type, specialist type, procedural stage, and verifier requirements. The context manifest records which records were loaded, why each was eligible, what was excluded, and which promotion or approval basis supports each loaded record.

The bounded-retrieval discipline parallels MemGPT's tiered memory management (Packer et al., 2023), which treats the model context window as a constrained resource that must be managed with explicit eligibility rules. The legal-workflow version replaces "what fits in the window" eligibility with "what is in scope for this matter under this jurisdiction" eligibility.

A firm can ask "what did this run consider, and why was each record eligible?" and get a structured answer pointing back to the manifest. A vendor that cannot answer that question does not have a memory architecture (FlowCounsel, "Why Legal AI Needs Bounded Memory, Not Bigger Prompts," April 2026).

Stating non-goals explicitly is part of the architecture. The patterns excluded by design are the patterns most often conflated with memory in current vendor pitches.

A legal AI product that claims "your firm gets smarter from your work" without articulating which memory class, what governance, what retrieval boundary, and what promotion discipline is making a marketing claim, not an architectural one. The two are distinguishable on a single question: can the vendor produce, on demand, a structured record of what the system considered on a specific run and why each record was eligible.

FirmIQ™ exists because that question has an answer architecturally, not just policy-wise. Memory architecture is governance applied to learning.

The structural advantage compounds over time as approved work accumulates and promotion runs through real attorney review. The marketing-label version does not compound. It conflates storage with learning, and the storage choice does not become more useful as the firm uses the system.