Recall recorded relationships

Prerequisites

Start ./scripts/dev.sh, open http://127.0.0.1:8787, and select a Brain you can read. Its selected knowledge or repository graph must be ready. See graph projection and combined views for rebuild and exact-manifest selection. Exact/text plus graph recall needs no model call. Semantic additions require the standing model and embedding policy.

Desktop procedure

In Recall memory, expand Scope, time and exact lookup and select graph alongside exact, text or semantic search. Graph discovers recorded neighbors of up to three qualified query matches; graph-only search is unavailable. Choose knowledge, repository or combined graph, direction, one to three hops and any relationship filters. The graph inherits the same Brain, collection, scope, manifest, fact time and eligibility mode as the rest of the query.

Repository graph requires one selected repository. With an environment, select its exact manifest; without one, the latest snapshot known when the request starts applies. Combined graph requires both environment and manifest. Historical mode or an explicit knowledge timestamp cannot be combined with graph yet. The UI explains these combinations before submission.

Keep source-coverage preference enabled to reserve some context for additional canonical source groups, then fill with ranked depth. It does not promise independent corroboration or impose a one-result-per-file limit. Results show the number of resolved groups and unknown-lineage items. Disabling preference allows an explicit comparison using the same query and budget.

Press Recall. Graph results show the anchor, recorded path and relationship families beside the result's own evidence links. A path explains discovery; it is not supporting evidence, acceptance, deployment state or permission to execute a procedure. Inspect evidence opens canonical retained evidence. Use the Brain's graph explorer for the wider bounded graph or a path query.

The full attributed context, including witnesses and provenance, must fit the selected byte budget. A large witness may leave fewer results; increase the explicit budget or narrow the query. Copying context includes its data-only instruction. Discard any external copy when scope or authority changes.

Scope/filter changes, graph invalidation and expiry clear displayed context. Failed refresh also removes prior results. These events do not silently repeat a paid semantic query; submit a fresh query explicitly when appropriate.

Native procedure

Pair the native companion and begin a context operation using workspace scope. Use the returned immutable operation ID:

./target/debug/recollect-agent scope recall "$BRAIN_ID" "$OPERATION_ID" "Vault JWT" \
  --channels exact,lexical,graph --graph-kind knowledge \
  --graph-direction both --graph-hops 2 --source-diversity true

--manifest ID supplies the parent exact manifest; --graph-relations accepts a comma-separated list of supported relation names. Repository/environment selection comes from the bound operation, even after the task default changes. Begin a new operation when that selection should change. Adding semantic creates one fresh query attempt per explicit invocation through the same API handler.

Failure and limits

• No eligible anchors: check exact identity, query text and scope; a safe recalled fragment cannot seed a whole-source vertex withheld by current qualification.
• No neighbors: a verified graph contains no connected qualified candidate under these direction/relation/hop settings. This is not proof of real-world absence.
• Missing/stale/damaged graph: inspect graph status and rebuild through its owner. A requested graph channel failure refuses the whole query without a silent fallback. Physical availability is checked before a paid query embedding.
• Partial graph: coverage names omitted projection inputs or the candidate bound (up to 100 discovered identities). Narrow scope/query; paths never widen scope.
• Input/capacity/timeout error: complete input is bounded to 5,000 vertices and 20,000 edges, with two graph read slots and a cumulative 15-second graph-work budget. Total recall allows 30 seconds with graph or 90 with semantic. Provider wait does not consume graph-work time; no automatic model retry is added.
• Insufficient support: inspect claim qualifications or use the strict modes. Semantic similarity and connectivity do not establish support for an answer. Unsupported synthetic questions still returned unrelated semantic context in the measured comparison.

Verification

With the repository-owned services available:

set -a
source .env
set +a
cargo test -p recollect-server --test platform graph::retrieval:: -- --ignored --test-threads=1
cargo test -p recollect-server --test platform retrieval::context:: -- --ignored --test-threads=1
./scripts/test-ui.sh tests/recall-graph.spec.ts
./scripts/validate.sh

These fixtures use synthetic inputs and controlled providers; no actual OpenAI charge is required. The actual-model evaluation script is a separate, explicitly isolated proof with retained attempt state, not a routine test or an automatic replay procedure. The mapping records its measurements and the normal SWEG runtime preservation check. Current UI verification targets desktop only.