feat(graph): report edge counts, and narrow the NREM read in SQL

[Cipher208]

What this fixes

memory_stats and the dashboard reported graph_nodes and nothing else. The size of the graph was invisible to every tool that could read it — the numbers behind this session's measurements were all raw SQL against the database files, which makes them incomparable between runs.

Changes

1. Edges become visible. New EpistemicGraph.count_edges() returns {"total", "live", "dead"} in one query, surfaced as graph_edges / graph_edges_live / graph_edges_dead in memory_stats and dashboard.get_stats. All three default to 0, so existing consumers are unaffected.

epi_edges has no layer column, so an edge is attributed to its source node's layer via a join — the convention memory_graph_edges already uses. An edge whose source node is gone lands in no bucket rather than a guessed one.

"Dead" is weight <= 0, not weight < NREM_FLOOR: across all three house databases only 0.1% of edges fall strictly between zero and the 0.05 floor (103 of 186 523), so the coarser split is the accurate one. A dead edge is not garbage — it is the state the next miner pass revives it from, which is why nothing prunes zero eagerly.

2. _dream_nrem reads only what it can act on. It previously read every heuristic edge and decided age in Python, discarding most: on the busiest database, 155 486 rows read to act on a few thousand, and 0 rows for mimocode's over-30-day band. The two actionable age bands are now the query's bounds.

The comparison is the same one the loop made — created_at < now − 30 d is age_days > 30, and created_at >= now − 1 d is age_days <= 1, inclusive, as before.

Database	rows before	rows after	reduction
hermes	155 486	77 857	49.9 %
mimocode	132 068	42 280	68.0 %
cowagent	49 772	28 817	42.1 %

A NULL created_at is now skipped instead of raising float(None). No house database has one; "unknown age" is not a reason to decay, boost, or prune.

No index, deliberately

An index on created_at was considered and rejected on measurement: the filter still reads 42–68% of the table and the time fell with the row count (190 → 115 ms, 205 → 79 ms). An index at that selectivity would buy nothing while keeping its write cost.

Tests

• four on count_edges — live/dead split, empty graph, source-node layer attribution, orphan edge;
• one on NREM — a ten-day-old edge is left exactly as it was, which is what proves narrowing the read kept the behaviour;
• test_stats and test_dashboard assert total == live + dead.

Full suite: 2078 passed.

Cost

count_edges with the layer join, measured on the live databases: 97 ms on hermes (186 878 edges), 30 ms on cowagent. It runs only when the tool is called.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: Repository: Cipher208/a-memory/.coderabbit.yaml
• Review profile: CHILL
• Plan: Advanced
• Run ID: 2370ecc8-2ef0-4664-acd7-dea8f4c856de

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}