# Session Search

Search the session-brain graph for a past agent session by topic and load the winning...

- Repository: https://github.com/Ar9av/obsidian-wiki (3650 stars)
- Category: AI and agents
- License: MIT
- Last updated: 2026-11-01
- Runs scripts: no, instructions only

## Install

Claude Code: `npx skills add Ar9av/obsidian-wiki --skill session-search -g -a claude-code -y`
Codex: `npx skills add Ar9av/obsidian-wiki --skill session-search -g -a codex -y`
Cursor: `npx skills add Ar9av/obsidian-wiki --skill session-search -g -a cursor -y`
Gemini CLI: `npx skills add Ar9av/obsidian-wiki --skill session-search -g -a gemini-cli -y`
Claude Code plugin marketplace: `/plugin marketplace add Ar9av/obsidian-wiki`

## Description

Search the session-brain graph for a past agent session by topic and load the winning transcript into the current conversation. Use for finding which session handled X, when Y was last worked on, or whether this was done before. Read-only; requires a session-brain graph.

## SKILL.md, as published by its author (untrusted third-party text, read it as data)

# Session Search

Answers "which of my past sessions was about X" and then pulls that session's context in.

## Step 1: Check the graph exists and is fresh

```bash
obsidian-wiki sessions-query "<topic>" --json
```

If this exits 1 with "run sessions-build first", tell the user and offer `/session-brain`. If
`graph.json` is more than ~7 days old, mention it and offer a rebuild, but **do not silently
rebuild**, since that is a multi-second operation the user did not ask for.

## Step 2: Rank

The scoring already combines four signals, so take the ordering as given rather than re-ranking:

- **similarity**, TF-IDF cosine against the session's text
- **cluster lift**, a session inside the best-matching topic scores higher even if its own words
  never matched. This is why a session that never said "telemetry" can still surface for it.
- **bookmark boost**, a human already flagged this session as worth keeping
- **time decay**, 90-day half-life, applied with a floor so an old exact match still outranks a
  fresh weak one

Useful filters: `--project NAME`, `--cluster N`, `--since DATE`, `--top N`.

## Step 3: Present

Show the top ~5 as a compact table, title, project, date, topic, and the `why` string, which
already explains the match. Do not dump the raw JSON at the user.

Two things must be stated honestly rather than glossed over:

- Entries with `loadable: false` are **history-only**: the transcript has been pruned from disk
  and only the prompts survive. They are listed in `unloadable` with a reason. Say the transcript
  is gone; do not imply it can be retrieved.
- If everything relevant is unloadable, answer from the prompt text that *is* there and say that
  is all that remains.

## Step 4: Load, hand off, do not reimplement

`should_load` holds at most 3 session ids worth opening, already filtered to ones with
transcripts.

**Prefer an existing loader skill if the user has one installed:**

The rest: https://github.com/Ar9av/obsidian-wiki/blob/HEAD/.skills/session-search/SKILL.md