js-ai-agent

js-ai-agent

Stateful AI agent with per-session conversation memory persisted in C3. Each session accumulates its full message history; that history is replayed to Claude on every turn so the agent remembers what was said earlier in the conversation. Multiple sessions are fully isolated from each other.

Requires an Anthropic API key (sk-ant-...).

Setup

ribo db create agent
ribo deploy
URL=http://localhost:8080/<addr>
curl -X POST $URL/init
curl -X POST $URL/configure -d '{"api_key":"sk-ant-..."}'

Optionally set a custom system prompt:

curl -X POST $URL/configure \
     -d '{"api_key":"sk-ant-...","system_prompt":"You are a pirate. Speak accordingly."}'

Routes

Method	Path	Body	Description
POST	/init	—	Create schema (sessions, messages, config)
POST	/configure	{ api_key, system_prompt? }	Store config in C3
POST	/sessions	—	Create a new session → { session_id }
GET	/sessions	—	List all sessions with turn counts
POST	/sessions/:id/message	{ content }	Send a message, get AI reply
GET	/sessions/:id/history	—	Full conversation history for a session
DELETE	/sessions/:id	—	Clear a session's messages (keeps the session)

Try it

URL=http://localhost:8080/<addr>

# start a session
SESSION=$(curl -s -X POST $URL/sessions | jq -r .session_id)

# multi-turn conversation — agent remembers earlier turns
curl -s -X POST $URL/sessions/$SESSION/message \
     -d '{"content":"My name is Alice. Remember that."}' | jq .reply

curl -s -X POST $URL/sessions/$SESSION/message \
     -d '{"content":"What is my name?"}' | jq .reply

# inspect history
curl $URL/sessions/$SESSION/history | jq .

# start a second, isolated session
SESSION2=$(curl -s -X POST $URL/sessions | jq -r .session_id)
curl -s -X POST $URL/sessions/$SESSION2/message \
     -d '{"content":"What is my name?"}' | jq .reply
# → agent has no idea — different session, no memory of Alice

Schema

config    (key TEXT PK, value TEXT)
sessions  (id TEXT PK, created_at, updated_at)
messages  (id INTEGER PK, session_id, role CHECK('user'|'assistant'), content, created_at)

What this demonstrates

• Stateful serverless: the Cell itself is stateless (fresh isolate per request), but state lives in C3 — giving you durable, queryable memory for free
• Conversation replay: history loaded from C3 and sent to the API on every turn; no in-memory accumulation, no sticky sessions, no servers to keep warm
• Session isolation: each session is a separate DB partition; concurrent users cannot see each other's conversations
• C3 as agent memory: the same pattern scales to tool call logs, embeddings, retrieved documents, or any other persistent agent state