session-brain indexes your Claude session history into a graph of topics. The heavy work is deterministic Python in the obsidian-wiki command line tool: it reads transcripts, links similar sessions and groups them into clusters. The agent only adds a short name and one-sentence summary for each cluster that has none yet.
The output is a sidecar folder at ~/.claude/session-brain, including a graph.html you can open in a browser. It reports the biggest topics, what is heating up, what has gone quiet and which sessions bridge two topics.
What You Get
- A topic graph of your session history stored in a sidecar folder, with a graph.html viewer
- Short names and summaries for every cluster, kept across rebuilds
- A report of the biggest topics, the hottest momentum, dormant topics and bridge sessions
- Tuning flags for stricter edges, mutual-neighbour clusters, recency half-life and per-project exclusion
- A clear count of sessions that have transcripts and sessions that only survive as prompts
Common Use Cases
- Getting a map of what you have worked on across hundreds of agent sessions
- Spotting topics that have gone quiet for two months or more
- Finding sessions that connect two otherwise separate topics
What Makes It Unique
Naming costs one turn whatever the size of the history, because the tool hands over top terms and exemplar titles and the agent never opens a transcript for it. The viewer encodes session length as node size, recency as brightness, history-only sessions as hollow rings and bookmarked sessions as gold borders, with a time slider and search box. It is explicitly a sidecar and does not touch the vault.
Under The Hood
The skill is instructions that run the obsidian-wiki command line tool through the shell: sessions-build, sessions-clusters and sessions-name, with names passed in through a heredoc. The tool reads Claude session transcripts and the prompt history file on disk and writes graph, cluster and name files under ~/.claude/session-brain. The skill also runs the open command to show graph.html in a browser. It does not fetch web pages or call remote APIs, and it does not write to the wiki vault.