Sessions

Every run writes a transcript — prompts, tool calls, responses — to a local store, indexed by content and project. agents sessions searches it across Claude, Codex, Gemini, and OpenCode, replays it, and streams it live.

Search and read

Pass a query to search this project by topic, file path, or command; pass a session id to open one. A filter flag without --markdown/--json defaults to markdown output.

agents sessions "stripe webhook signature"       # search this project
agents sessions "deploy script" --all            # search every directory
agents sessions a7f3e2c1 --markdown              # replay as markdown
agents sessions a7f3e2c1 --include user          # just the user turns

--markdown renders user, assistant, thinking, and tool calls with secrets redacted by default. Narrow the replay with --include/--exclude <roles> or --first/--last <n> turns. The search reaches across every agent — one query, every CLI's history.

Live sessions

--active shows what is running right now across terminals, teams, cloud, and headless runs, grouped by directory. Each row carries a live state:

• working — the agent is mid-turn, executing tools or generating.
• waiting — it is blocked on your input. --waiting filters to just these (and exits non-zero if any exist, so a shell prompt can nudge you).
• idle — the session is open but the agent is not doing anything.

agents sessions --active           # everything running now, by directory
agents sessions --active --waiting # only sessions blocked on your input

Focus a live session

sessions focus jumps you into a running session — it attaches the terminal the session lives in, or opens a new tab and resumes it. Omit the id for an interactive picker.

agents sessions focus 80289224      # attach or resume a running session
agents sessions focus --attach-only # attach only, never open a new tab

Stream events

sessions tail follows a session's JSONL events as they are written — one event per line, Claude and Codex only. Start at the end, or replay from the beginning first.

agents sessions tail --latest                 # follow the most recent session
agents sessions tail a7f3e2c1 --from-start    # replay, then follow
agents sessions tail --latest | jq 'select(.type == "user")'

Query another machine

--host <name> runs the query on a remote machine's own index over SSH — no sync needed, SSH access is the only auth. Repeat it to fan the same query across several machines. The interactive listing folds in your other online machines automatically; --local keeps it to this box.

agents sessions "auth bug" --host yosemite-s1        # one remote index
agents sessions "deploy script" --host box-a --host box-b   # fan out
agents sessions --active --local                     # this machine only

Across devices

History also syncs across the fleet via CRDT merge, so a session started on one machine is readable everywhere. Sessions across devices →

How it works — index, not embeddings

"Agent memory" usually means a vector database: embed every message, hope the nearest neighbor is relevant. Sessions takes the opposite bet. The transcript each agent already writes is the source of truth — agi-cli indexes that text into SQLite full-text search (FTS5, BM25 ranking), with no embeddings anywhere.

Each search compares every transcript file's modified-time and size against a ledger and re-parses only what changed, so the index is current the instant a file is written and a search is a local index lookup — no network call, no model, no re-encode tax. That is why recall is instant, always fresh, and free.

To steer a running agent rather than read a finished one, see Messaging.

Related: Learn: How agents communicate, Messaging, Teams, Cloud dispatch, Sessions across devices.