Agent memory

Every coding agent session starts from scratch unless something carries context forward. The memory snippet gives agents that something: commands to search all past sessions, read any conversation, watch live ones, and pull relevant prior work into the current task. Memory works across tools — a Claude Code session can recall what was built in Cursor — and across the team, so an agent can learn from a colleague's session as easily as its own.

Memory is the foundation snippet: installing it via the snippet system also enables messaging by default, and the tasks and plans snippet rides along with it.

cast memory        # install
cast install memory --disable

Search and browse

The snippet teaches agents to search proactively — when starting a task, when debugging, when the user references previous work — and to parallelize searches across topics. Default scope is the team, from the current directory's project:

cast search "auth"                # team-wide search
cast search "auth" --mine         # only my sessions
cast search "auth" -m samvit      # a specific member
cast search "auth" -g -s 7d       # all teams, last 7 days
cast feed                         # recent team sessions
cast feed --state needs-input     # filter by work state
cast read jx7c6zk 15:25           # read messages 15–25 of a session
cast read '<share-url>#msg-<id>'  # read a window around a linked message
cast link jx7c6zk 42              # mint a deep link to any object

Search is hybrid: keyword matching combined with semantic similarity, with a fallback to title search under load. Results return original conversation fragments, not summaries, so the agent gets precise, quotable context.

Watching sessions live

cast sessions is the state axis of memory — not what was said, but where every session stands right now:

cast sessions                     # snapshot, grouped most actionable first
cast sessions -w                  # live stream: one line per work state change
cast sessions -w --json           # …as NDJSON events for scripting
cast sessions --label fleet -w    # watch every session filed under a label
cast sessions --messages -w       # follow messages across live sessions

The -w stream is silent until something changes, which makes it a reliable wake signal for orchestration loops (see messaging for the spawn → watch → send pattern). needs_input means the ball is in your court — the session finished its turn, asked a question, or hit a permission prompt.

Labels are personal filing: cast label set api jx7c6zk files a session under a name, and every browse command takes --label to filter by it.

Analysis and continuity

cast diff jx7c6zk                 # files changed, commits, tools used
cast diff --today                 # aggregate today's work
cast summary jx7c6zk              # goal, approach, outcome, files
cast context "implement auth"     # find relevant prior sessions before starting
cast ask "how does session sync work"   # natural language answer across sessions
cast handoff                      # generate a context transfer document
cast decisions add "Use Stripe Checkout" --reason "handles SCA"
cast decisions list

cast context is the habit that pays off most: run it before starting anything and the agent begins with the three most relevant prior sessions instead of rediscovering them. cast handoff closes the other end — it distills the current session's goal, approach, and open items into a document the next session can start from.

Ambient recall

Search is recall on demand. Ambient awareness is recall pushed: with stable mode on, every new session starts with a feed of recent sessions already in context, no search required. The two compose — the feed tells the agent which sessions exist and their state; cast read and cast search go deep on the ones that matter.