CoCounsel Agentic Memory

Problem

Legal practitioners face unreliable, un-auditable AI memory when using general-purpose assistants for matter work. Lawyers repeatedly re-enter matter context, risk reliance on stale or overruled law, and face potential ethical-wall violations across matters. That erodes trust in AI-assisted work product, increases malpractice and confidentiality exposure (as seen in United States v. Heppner), and lets institutional expertise walk out the door when senior practitioners leave—rather than being captured as reusable, auditable knowledge.

Constraints

Memory for legal AI has to meet a fiduciary bar: staleness, ethical walls, override, and opaque capture each need a concrete architectural answer. Grounding must resolve to owned canonical authority (Westlaw, Practical Law, KeyCite)—not partner integrations or MCP calls that can't reach underlying taxonomy. Ethical walls must be deterministic: matter scope and access boundaries stamped at write time and recomputed on every read, so the model is never the last line of defense.

Process

I produced 18 design variations exploring how memory, grounding, and precedence surface across interaction models; authored a project blueprint for phased scope across semantic memory, procedural memory, and grounding; and ran competitive analysis of agentic memory approaches (bi-temporal graphs, schema stores, vector recall with recency decay, hybrids) to find where legal requirements diverge. Competitive analysis artifacts will be added here in a later upload. I partnered with Product to sequence MVP capabilities—preference memory and precedence first versus procedural promotion and cross-matter sharing—and opened a structured feedback loop with KMs, PSLs, and legal ops on unresolved architecture questions before locking the system.

The turning-point decision

The turning point was separating two memory types legal work actually needs. Semantic memory covers preferences, tone, and playbooks. Procedural memory covers how work is actually done—and what happened after—the harder, higher-value problem nobody had built yet. Governance for preference promotion and precedence sits with KM, PSL, and legal ops—the functions that already govern legal knowledge—not with the model.

Final solution

Semantic vs. procedural: Treat preferences and playbooks as a shippable layer first; design procedural memory as a two-step pipeline where individual episodes consolidate into reinforced how-to patterns, then promote into versioned, human-reviewed skill documents. Ground to owned authority: Resolve extracted text and citations to Westlaw, Practical Law, and KeyCite so answers stay auditable—and so classification can realistically map entities to canonical authority, validated against gold-standard documents via Anthropic API metadata pulls. Write-time ethical walls: Compute and stamp matter scope at write time; recompute on every read. Remove reliance on the model as the confidentiality backstop. Measure before asserting: An evaluation framework—recall, precision, and honest abstention—lets the team test memory claims before launch instead of hoping they hold.

Measured outcome

We defined and validated a fiduciary-duty framework mapping each general-memory failure mode to an architectural answer; produced a testable evaluation framework ahead of launch; achieved cross-functional alignment on separating semantic from procedural memory for MVP scope; and opened a practitioner feedback loop before finalizing architecture. Full metrics collection (adoption, task success, drafting quality) is planned post-launch once grounded memory and deterministic-wall architecture ship.

What I learned

For legal AI, memory is not a convenience feature—it is a trust and ethics surface. Making capture, precedence, and walls legible to practitioners mattered as much as the underlying store. The work that shipped first was the work we could prove: preferences and governance people already understand, with procedural memory designed as the harder next phase rather than a vague promise.