AF-Memory — product definition

The sentence: Connect it to your document mess. Ask questions. Get correct, cited answers. It checks itself.

Date: July 22, 2026. Status: definition approved direction; ask()/sync() implementation in progress on feat/memory.

What the customer sees — five verbs, nothing else

Verb	Customer experience
connect	Pick the sources: a SharePoint library, a shared inbox, a folder. That choice is the governance act.
sync	"631 documents in memory, current as of 9:02." Runs continuously; backfill is an afternoon.
ask	A question in plain language; an answer with citations to the source documents — or an honest "that isn't in your documents."
check	A trust report card after every sync: "31 of 32 checks passed; 1 fact needs review."
pin	One click on any human-verified fact makes it a permanent check that every future sync must pass.

What is deliberately hidden (and must stay hidden)

• Engine names and engine choice. One product. Any internal index (graph, vector, or otherwise) is an implementation detail the router uses; the customer never selects a backend.
• Query strategy. Measured fact from our benchmarks: answer quality depends on how the store is queried, and exposing raw tools degrades results (an agent given files and graph together scored below files alone). The ask() router owns strategy: attribute questions go to entity profiles, counts to deterministic scans, multi-hop questions to traversal, document lookups to source text — and it abstains rather than guesses.
• Ingestion mechanics. Retries, resume, skip-and-report on pathological documents, model-response repair — all built, all invisible. The customer sees a document count and a trust score.
• Tuning surface. Workspaces are created automatically per program (no "group IDs"). Ontologies are optional domain packs with a sane default. Temperatures, embedding dimensions, prompt hints: never customer-visible.
• The agent tool surface. Agents get two tools — ask and lookup — with routing built in. The raw primitives remain behind a power-user flag.

Why the hidden parts exist (internal, for engineering)

Every hidden mechanism is receipted by a measurement from the July 2026 benchmark and demo builds: routing (hybrid-tool degradation, 42/50 vs 50/50), abstention (confidently-wrong counts dominate errors in every naive mode), deterministic counting (ranked search miscounts; scan doesn't), temporal reconstruction (amendment-chain state), ingest resilience (deterministic pathological documents in 2 of 3 corpora), and the verification gate (rejected multiple bad builds during a single week). See client-goa/demo/specs-cognee/COMPARISON.md and the demo battery receipts.

Build sequence

1. Memory.ask() — the routed answerer — and Memory.sync() with built-in invariant checks (in progress, feat/memory).
2. AF-Memory MCP server collapses to ask + lookup (primitives behind a flag) — after (1) lands.
3. Control Centre panel: Ask box, Trust tab with pin-on-verify.
4. Connect-source wizard (SharePoint/Outlook via Microsoft Graph API) — the last mile that removes the folder assumption.
5. Internal second index behind the router where measurements justify it. Never a customer-facing choice.