vesseld with conversational history

Single-vessel, single-agent demo that wires a shared HistoryStore so the agent remembers what was said earlier in the conversation. Pair this with vesseld-multi-vessel — that one shows multi-agent + Kanban delegation; this one shows how a Vessel preserves transcript state across turns.

What it shows

File	Concept
daemon.yaml	Daemon-wide config: control socket, drain timeout, shared LLM rate-limit bucket.
shared/openai.yaml	One LLMProfile.
shared/history.yaml	A HistoryStore with ref: buffer and a per-conversation cap of 500 messages.
vessels/assistant/vessel.yaml	Vessel references the HistoryStore; the Agent opts into historyAccess: read_write.

Multiple Vessels can reference the same HistoryStore — useful when a "human" vessel and a "supervisor sidecar" vessel need to see one another's turns.

Prerequisites

export OPENAI_API_KEY=sk-...
go build -o ./vesseld ./cmd/vesseld

Run

./vesseld validate --config examples/vesseld-with-history -R
./vesseld run      --config examples/vesseld-with-history -R

Drive a multi-turn conversation

The key is a stable context_id in each request. Different context_ids give independent transcripts; the same one continues a thread.

SOCK=/tmp/vesseld-with-history.sock
CONV=demo-alice                     # any stable string; treat it as the conversation key

# Turn 1: tell the agent something.
curl --unix-socket $SOCK -X POST http://vesseld/v1/vessels/assistant/call \
  -H 'content-type: application/json' \
  -d "{\"agent\":\"assistant-agent\",\"context_id\":\"$CONV\",\"query\":\"My name is Alice and I prefer coffee over tea.\"}"

# Turn 2: ask a follow-up — the agent recalls from the buffer.
curl --unix-socket $SOCK -X POST http://vesseld/v1/vessels/assistant/call \
  -H 'content-type: application/json' \
  -d "{\"agent\":\"assistant-agent\",\"context_id\":\"$CONV\",\"query\":\"What should I drink this morning?\"}"

The second response should reference coffee. Change CONV to a new value and ask the same question — the agent will (correctly) admit it doesn't know.

Caveats — what this is NOT

The buffer history lives in the daemon's memory. Restart vesseld and every conversation is gone.

Persistent transcripts across daemon restarts arrive in v0.2.0 via the Compacted history flavor (hierarchical SummaryDAG + SQLite/Postgres-backed Store). The declarative YAML schema and the vessel.WithSessionStore option are on the v0.2.0 roadmap.

If you need cross-restart memory today, embed vessel as a library and pass history.NewCompacted(store, llm, ws) directly — see examples/chatbot-with-recall/ for the in-process pattern.

Going further

• Set maxMessages lower (e.g. 20) and observe the agent "forgetting" once eviction kicks in.
• Add a second agent to the same vessel with historyAccess: read_only to model a supervisor that watches without writing.
• Combine with examples/vesseld-multi-vessel/ patterns: point both demo daemons at the same HistoryStore document and they'll share transcripts.