Session Brain

Builds a searchable topic graph over your agent session history. The output is a sidecar at ~/.claude/session-brain/, the vault is never touched.

All the heavy lifting is deterministic Python in the obsidian-wiki CLI. Your only job is to name the clusters, which takes exactly one turn and requires reading no transcripts.

When to use which skill

Goal	Skill
Build or refresh the graph; survey topics	session-brain (this one)
Find and load a specific past session	session-search
Distil sessions into permanent vault pages	wiki-history-ingest / claude-history-ingest

Step 1: Build

obsidian-wiki sessions-build --json

Roughly 3 seconds cold on ~1000 sessions, well under a second incrementally, it re-reads only transcripts whose size or mtime changed. Useful flags:

Flag	When
--full	Ignore all caches and re-read everything
--mutual	Tighter, smaller clusters (mutual-kNN edges only)
--half-life N	Change the recency half-life (default 90 days)
--min-sim 0.15	Fewer, stronger edges, use if the graph is too dense to read
--skip name	Exclude a project. Match is substring-based; pass the bare name, because cache dirs start with - and argparse reads that as a flag

Report the headline numbers: total sessions, how many have transcripts vs. are history-only, edges, and cluster count.

Step 2: Name the unnamed clusters

obsidian-wiki sessions-clusters --unnamed --json

Each cluster comes with top_terms and exemplars (its three highest-degree sessions, whose titles are already in graph.json). That is all you need. Do not open transcripts to name a cluster, the whole design goal is that naming costs one turn regardless of corpus size.

Write a 3, 5 word name and a one-sentence summary per cluster, then: