Recall

An agent that loads a lean bundle at session start still has to find the right document, fact, or task in the middle of a job. WilsonOS answers that with structure first and search second — a pattern the team calls retrieval-light.

How finding works

Layer	What it is
A manifest per brand	Every brand carries a navigation index listing each of its documents with a one-line role. A skill that needs brand context loads this first and follows it.
Links between documents	Brand documents cross-reference each other with wiki-style links. The knowledge browser renders them clickably; agents resolve them with the same path resolver.
A backlinks index	For any document, which documents refer to it — so an agent can discover related material before proposing something new, and so nobody writes a fifth document about a thing that already has four.
Keyword search	Over the whole corpus — tasks, skills, generated media, brands, memory — scoped to the searcher. The app’s universal search and the agent’s search command hit the same index.
Skills declare their context	Each skill names, up front, the documents it needs. The runner loads them; the agent does not have to discover anything at run time for routine work.

The last row is the quiet workhorse. Most retrieval in a well-designed skill is not retrieval at all — it is a declaration, made once by the skill’s author, of exactly what to read.

Search is scoped

A search carries the identity of whoever is searching. Brand-shared classes are visible to anyone with brand access; private memory rows appear only for their owner; older, unscoped notes appear only to an administrator during review. A search with no principal returns shared classes only and says so, rather than quietly widening.

Why not vector search (yet)

Dense-embedding retrieval is the default answer in most agent stacks. WilsonOS defers it on purpose:

• It has real cost (embedding calls, a vector store) and real maintenance (re-embed on every edit, drift detection, chunking and hybrid-scoring choices).
• The corpus is a wiki-shaped body of documents with manifests and links — structure that keyword search plus navigation already exploits well.
• The payoff over the current approach is unproven at this scale. It will be adopted when keyword-plus-structure demonstrably fails on real work, not before — and when it is, it will be an additional layer over the existing structure, not a replacement.

The empty-search rule

No matches is not proof of absence. Agents are required to verify the structure — the manifest, the path, the index — before concluding something does not exist, and never to treat a suppressed error as an empty result. It sounds small; it has prevented real wrong answers.

What recall is not

It is not the boot bundle (that is session start), and it is not the memory store (that is what your agent remembers). Recall is the path between them: how the right slice of what is stored reaches the agent at the moment it is needed.

live Brand manifests, wiki links and the backlinks index, scoped keyword search, and skill-declared context are live. proposed Semantic search is a direction to revisit after reliable sources and access are in place; nothing here commissions a vector service.

Related

What your agent remembersWhat there is to find.

Session startWhat is already loaded before recall is needed.

SkillsDeclaring context up front instead of searching for it.

Your brandsThe manifests and links live in the brand stack.