Go-to-Market Context Is a System, Not a Document

The unlock: context that persists

Every AI surface arrives amnesiac: it knows nothing until you feed it, and the copy you feed it starts going stale the moment you do. I ran that loop for two years at WorkSpan - n8n flows, Clay tables, Claude projects, the email stack. Everywhere we deployed AI, we had to deploy a lot of context, and it went stale quickly. A sales kickoff on new packaging would end with a spotty list of prompts and context stores to update by hand. One step forward, one step back. The burden has a name - context debt - and it compounds with every launch, every competitor, every re-cut ICP.

Field noteContext debt: the liability you're already carrying →The mechanics, the five symptoms, and the crossover where a context system turns the liability into an asset.

Watch what the debt actually looks like in motion - and what replaces it.

The write path is what made the new world real for us. We had said no to Octave, a semantic graph built for go-to-market context, a year before we adopted it - "there's no way I'm going to go fill out 50 different objects and write pages and pages of text" - and an MCP that let Claude write the graph flipped the answer: one working session and three days of kickoff transcripts got it "probably 60% built out." From then on, maintenance meant feeding it transcripts. A new play stopped costing a Monday enablement session; it was a Slack message - "new play dropped in Octave" - and reps learned the play by running it.

Brendan Short hosted the live demo of this system in April 2026 and named both states on the record. The before, from his own enablement years: "I got to go enable 100 reps across the org on this new motion or new positioning. Very painful, especially as you get bigger." The after, watching the layer update itself from a transcript: "Being able to have this living, breathing thing that can just auto-update is very, very crazy. Kind of breaks my brain."

Persistent does not mean always-loaded

Persistent context survives between sessions; always-loaded context rides in front of every prompt whether the task needs it or not, and it degrades the work. Chroma measured output reliability falling as input grows, across eighteen frontier models - the field calls it context rot - and Anthropic cut over 80 percent of Claude Code's system prompt with no measurable loss: "less always-loaded context, not less context." Both are coding results, and coding is the domain that gets its context for free. Go-to-market has to write its own - which makes the real question where the context lives and who owns it. That is an architecture decision.

Field noteInside the WorkSpan run: context that compounds instead of going stale →The full production case - what got maintained before and after, the propagation math at each surface, and the transcript-fed maintenance loop.

Context lives in layers

Committed teams then make the second mistake on schedule: treating everything the org knows as one kind of thing, in one store, on one refresh cycle. Four questions say otherwise - who owns it, how fast does it decay, is there a right answer, what happens if it is wrong - and asked across everything your team knows, the answers cluster into three layers.

By mid-2026 the shape stopped being a private position: Kieran Flanagan scopes "AI brains" personal, team, company; the playbooks PE operating partners hand their portfolio companies scope org, function, user; Gartner forecasts most agentic systems running on context graphs by 2028. The layer counts differ; the shape is the same. What none of them settle is who owns each layer, what may live where, and what keeps it fresh - the operating model, the half that does the work.

The operating system: what lives where, and what keeps it fresh

Running a context system comes down to three routing rules and five freshness mechanisms, all of them deliberately boring. Freshness that depends on heroics is decay on a delay.

Rule 1 - One fact, one home. Every fact is authoritative in exactly one layer; everything else links to it or retrieves it. Two copies is not redundancy, it is a race: one update, and every agent downstream inherits the contradiction.

Rule 2 - When it doesn't fit, ask the four questions. Some context refuses to bin: the quarter's OKRs, a project scratchpad. Don't invent a fourth layer; give that content its own small system with its own owner and clock. The layer model is a lens, not a periodic table.

Rule 3 - Owners the org already has. Context infrastructure dies when maintaining it is a job nobody was hired for. Each layer maps to a role that exists: product marketing already owns positioning, RevOps already owns the CRM, market intelligence already reads the market. Their outputs change shape, nobody takes a second job, and no single resignation kills the brain. The same test governs tooling - not which tool scores highest, but which system your people can keep alive. Fit to the org, fit to the people, then quality and cost.

The five freshness mechanisms

Ownership decides who. These decide how it stays true, in order of leverage.

I. Write at the point of change. Feed the transcript in the meeting where the change happened; approve the write. A separate maintenance project loses to the quarter every time; a ninety-second habit wins.

II. Match each clock to its decay rate. Conceptual: continuous intake, weekly curation. Canonical: editorial review, bi-weekly to monthly. Deterministic: sync-driven, engineering SLAs. One cadence for everything is wrong twice over.

III. Curate with verdicts. Every incoming signal gets one of four: accept whole, attach as evidence, rewrite into the house position, reject and log why. The logged rejections are how the bar teaches itself to the next curator, human or agent.

IV. Run canaries. Decay is invisible from inside; outputs just get quietly blander. Re-run the same account plan on a cadence - when it comes back vaguer than last month's, a layer went stale, and the diff says which.

V. Give agents memory with receipts. Hightouch's RevOps agents record what they recommended in version-controlled files and read it back next run - "it reads the recommendation it gave last time and factors that in." An agent that remembers coaches; one that doesn't nags. Files the team owns also mean you can swap execution tools without rebuilding the intelligence.

Field noteThe benchmark: six context architectures, same tasks, blind grading →The proof under this piece - context shape beats context volume, with prompts, traces, outputs, and rubric published so the result can be checked rather than asserted.