Investigate recalled memory on desktop

For the current route/menu map, see the desktop guide.

Prerequisites

Run ./scripts/dev.sh, open http://127.0.0.1:8787, sign in and select a Brain. Start with canonical recall; optional semantic and graph recall use the same selection and lifecycle. Investigation does not require manual review of normal learning or maintenance. Current UI work targets desktop.

Search and compare

Open Ask → Search evidence and enter a query or exact identity. Select an eligibility mode, collection, scope, optional exact manifest and time filters. Maximum results accepts one to twenty; the context budget includes full provenance and graph witnesses. Press Recall explicitly. The selection card records the actual Brain, named scope, manifest and server knowledge cutoff used for this response.

Use the three views without running another query:

• Matches shows the returned records, separate trust/lifecycle states, applicability, qualifications, channels and exact source locations. A policy acceptance is labeled separately from an authenticated account's review.
• Disagreements compares canonical conflict links from this result set. A related claim outside the bounded response has an explicit inspection link; opening it does not add it to context or claim it matched the query. Ten links are displayed per page. An empty view does not establish global consistency.
• Sources groups exact evidence references cited by the returned records, preserving used spans and repository commits. Shared citations do not establish independent corroboration. Missing bytes remain explicitly unavailable.

Copy attributed context copies the original bounded response regardless of the active view. It does not copy extra inspected claims or sources. Clipboard failure is displayed. External copies cannot be revoked by the browser; discard them when scope or canonical authority changes.

Inspect history, evidence and optional review

Inspect claim and history starts at the query's frozen knowledge cutoff and fact-time selection. A claim revised after recall therefore opens the version that the query saw. Choose Latest knowledge explicitly to inspect newer information. Prior revisions, unknown validity, observation precision, review, freshness and operational assessment remain distinct.

Procedure details include conditions, ordered steps and actual outcomes; handovers retain contributor inspection. No execution is authorized by opening or recalling them. Review and corrections uses the existing optional memory workflow. Readers and archived Brains cannot mutate, and historical revisions require an explicit return to current knowledge. A successful correction clears the investigation and sibling views; submit a new query to see its effect.

Inspect evidence and Inspect exact source load the precise retained source version or repository fact. They do not substitute newer bytes. A retained claim may outlive its raw support; the evidence dialog explains missing/expired content. Transient read failures hide the previous body and offer Refresh evidence.

Follow recorded relationships

On a current graph-compatible result, choose Explore relationships. This opens the existing native explorer around its exact revision key, even if it is outside the graph's first entity page. The dialog preserves the parent scope, collection, fact-time mode, exact manifest and returned repository snapshot. Plain exact/text results can open this view without a query embedding.

The graph is a fresh current read, labeled separately from the recall cutoff. Historical selections explain why graph exploration is unavailable. Missing or withheld centers, incomplete projections and scope limits display errors rather than switching to a broader graph.

Expand Exploration settings to change direction/hops or request an eligible overview. Select nodes/edges to inspect evidence and choose path endpoints. Shortest path controls uses the same native path handler; successful paths can be displayed on the canvas. Connectivity explains recorded relationships, not runtime impact or truth. Existing graph bounds and readiness procedures are documented in desktop graph exploration.

Clearing and recovery

Filter/Brain changes discard the prior response. Local mutations, observed permission changes, advertised deadlines and remote eligibility changes clear all investigation views and nested dialogs. The server metadata poll runs every three seconds while results are displayed; remote changes are observed on that bounded schedule, not through a push guarantee. No invalidation silently repeats a paid semantic query.

If Investigation cleared appears, inspect the current selection and submit a fresh query when appropriate. No-match, insufficient eligible evidence and partial coverage have separate messages. A high similarity or a graph path does not establish enough evidence to answer an unsupported question.

Verification

With repository-owned services running:

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

The fixtures are synthetic and require no actual model charge. The dated evidence records real database/native/browser calls, controlled browser error/deadline cases and the normal retained-runtime check. Changes remain local and uncommitted.