Client/server with Wingman
Wingman and Acumen are not two separate products. They are two halves of one product, with the relationship being the architecture itself.
Wingman is the client. Per-user state. Per-device deployment. The chat surface, local conversation context, user-specific preferences. Wingman knows the user.
Acumen is the server. Shared organisational state. Encoded memory, skill effectiveness measurements, learned practice accumulation, consolidated cross-user context, system-hole detection. Acumen knows the business.
Encoding-indexed memory
Acumen is at its core an encoding-indexed database with intention-based encoding. Not a relational store with vector retrieval bolted on; the encoding is the indexing. The data structure is organised around encoded representations of meaning, and retrieval happens through encoding similarity rather than through structured queries against typed records.
Intention-based encoding is the specific commitment that distinguishes Acumen from a generic semantic-embedding store. Generic semantic embedding captures what text means. Intention-based encoding captures what the speaker was trying to do. A user asking a question, a decision being made, a frustration being expressed, a workaround being applied — these encode based on their intentional shape, not just their topical content.
The four functions fall out naturally
- Cross-user consolidationThe act of placing every interaction into the encoding-indexed memory. The encoding does the structural work; no extraction-into-schema step required.
- Context-boostingEncoding-similarity retrieval against the user's current request. Content with similar intentional shape from prior interactions surfaces as relevant context.
- Cross-user suggestions and system-hole detectionEncoding-similarity matches across users. Cluster density above a noise threshold is the system-hole signal. No explicit threshold counting required.
- AI-interaction-pattern observationThe same clustering mechanism applied to skilled corrections of AI failure modes. Successful intervention patterns cluster as candidate Wingman skills.
Self-generated interfaces to sources of truth
Acumen doesn't pre-build connectors to organisational systems. It drives the generation of interfaces through Forge as the need surfaces.
When Acumen recognises that a source of truth must be reached — a CRM containing customer history, an ERP holding inventory state, a SharePoint instance with policy documents, a custom internal database — and no interface to that source yet exists, Acumen drives Forge to build the interface. Forge generates the connector as a standard build output. Guardian enforces the connector at runtime. Acumen now has access.
Custom organisational systems aren't a barrier. Most enterprise software platforms struggle with organisations whose internal systems don't have off-the-shelf connectors. Acumen handles them the same way it handles standard systems — drive Forge, build the interface, run under Guardian.
Compounding properties
Acumen's value compounds in distinct dimensions simultaneously, each reinforcing the others.
Institutional memory accumulation. Year one of a deployment, Acumen is thin. Year three, Acumen holds substantial portions of the organisation's tacit knowledge in queryable form.
Validated infrastructure compounding. Through the Acumen → Forge → Guardian system-hole loop, missing infrastructure gets built specifically for the deploying organisation. Each system hole detected becomes a bespoke module that fits the organisation's actual workflow.
Connector library compounding. Generated interfaces accumulate as a deployment-specific connector library. Generated patterns for common systems flow back into Forge's reusable substrate.
Skill library compounding. Through the AI-interaction-pattern function, the Wingman skill library grows over time. Effective patterns developed by skilled users get harvested into shared skills available to everyone.
The commercial moat
The accumulated context in Acumen is not portable, not replicable, not abandonable without leaving institutional memory behind. This is not vendor lock-in by contract or technical obstruction. It's lock-in by accumulated value. Compounds with every passing quarter.
Status
Architecture commitments stable: encoding-indexed memory with intention-based encoding; self-generated interfaces via Forge; the four functions; client/server relationship with Wingman; closed loops to Forge and Guardian. Specific implementation details (encoding model, vector store substrate, privacy/access-control model, interface generation triggering) in formation.