Dashboard MEMORIES tile counts only memory_save facts, not the semantic/summary corpus (reads "1" on a healthy install)

[janacm]

Summary

On the viewer Dashboard, the MEMORIES stat tile counts only the discrete memory_save "fact" namespace (mem:memories), not the long-term memory corpus that auto-capture actually produces (semantic memories, session summaries, lessons, insights, crystals). On a healthy install with thousands of captured items, the tile reads MEMORIES: 1, which strongly reads as "memory capture is broken" even though everything is working.

This is a labeling / metric-scoping bug, not a data or endpoint failure. It is distinct from #544 (closed), where full-list endpoints 500'd on a large corpus and the viewer fell back to 0. Here the endpoints succeed and every other tile is accurate — the count is simply scoped to the wrong store.

Environment

• agentmemory plugin (Claude Code): 0.9.26
• runtime @agentmemory/agentmemory: 0.9.27
• engine (iii): 0.11.2
• OS: macOS, native/local server, viewer at http://localhost:3113, daemon at :3111
• No AGENTMEMORY_SECRET

What the dashboard shows

MEMORIES
1
latest versions

While the same install, via memory_diagnose, reports a large, healthy corpus:

Store	Count
Semantic memories	1,640
Session summaries	79
Lessons	899
Crystals	163
Insights	640
memory_save facts	1 ← what the tile counts

The one "fact" is a manual test write ("Test write to verify agentmemory is functional"), so the tile happens to render 1.

Evidence that capture is healthy

memory_diagnose → all stores pass ("All 1640 semantic memories are consistent", "All 79 session summaries are consistent", "All 899 lessons are healthy", etc.).

The compress pipeline is actively running on real sessions, e.g. audit entries:

• mem::summarize — observationCount 822 — "Split-button reset UX with undo toast and dev database branching"
• mem::summarize — observationCount 365 — "Refactor, test suite, and production bug fix"
• mem::summarize — observationCount 67 — "Fix /scheduled-tasks 404 and consolidate duplicate proxy files"
• mem::consolidate-pipeline — semantic.newFacts: 26, totalSummaries: 79

memory_recall on a recent topic returns real file_edit / file_write / command_run observations with session IDs and timestamps. So capture → compress → recall works end to end.

Expected behavior

The MEMORIES tile should reflect the long-term memory corpus a user thinks of as "memories" — i.e. include semantic memories / session summaries (and/or surface them as their own labeled tiles), rather than counting only the rarely-used discrete memory_save fact namespace.

At minimum, one of:

1. Have the tile count the semantic/summary corpus (what users mean by "memories"), or
2. Rename the existing tile to something accurate for that namespace (e.g. "Saved facts"), and add a separate "Semantic" / "Summaries" tile.

As-is, a fully-working install presents as empty on the primary dashboard, which reads as a capture failure.