Memory #

Track: Phase 1

Goal #

Create or open an agent that retains identity or conversation state across a disconnect / second turn, then show that prior context still affects the later reply.

Success criteria #

• Turn 1 establishes a memorable fact or persona.
• After ending the first connection (or starting a new turn on a durable id), turn 2 refers to that fact without re-sending the full history from the client.
• The second reply clearly depends on the retained state.

Assert (smoke) #

AGENT_LAB_RESULT {"v":1,"scenario":"memory","ok":true,"checks":{"durable_id":true,"fact_recalled":true,"reconnect":true}} Missing keys → AGENT_LAB_SKIP with scenario":"memory".

Check id	Required	Pass when
durable_id	yes	Non-empty durable session/agent/thread id used across turns
fact_recalled	yes	Turn-2 reply contains the planted fact (Orchid for the suggested prompt)
reconnect	yes	Connection/process truly ended before turn 2 (true only after dispose/exit/resume path)

Observe-only: richer “depends on state” beyond the fact substring (v1 uses fact_recalled as proxy).

Constraints #

• Persistence mechanism is up to the stack (server memory blocks, checkpointer + thread id, durable instance state, session resume, etc.).
• In-process variables that disappear when the script exits do not count.
• No Modal. Do not invent Decisions clients or undocumented APIs.
• Soft-pedal employer Managed platforms as first fills — prefer stacks with a documented durable session/thread/resume API.

Prompt (suggested) #

1. Remember that the project codename is Orchid.
2. (reconnect / new session on same agent) What is the project codename?

---

Mental model #

Memory is durable identity + true reconnect: turn 1 plants a fact under a stable session/agent/thread id; the process or connection ends; turn 2 resumes that id without the client replaying the full history, and the reply still contains the fact. In-process variables, module globals, or “keep the same messages[] array in RAM” die with the process and do not count — reconnect must reflect dispose/exit/resume. Failure looks like a two-turn chat in one process that never dropped the connection, or a client that re-sends the planted fact on turn 2.

What excellent looks like #

• Capture a durable id after turn 1; dispose or exit; resume by id for turn 2; assert Orchid in the second reply.
• Document the stack primitive in the folder README (conversation id, session.resume, checkpointer thread_id, Durable Object state, etc.).
• Preferred stacks when Eng fills: TypeScript — Letta resumeSession, openai-agents Conversations session, claude-agent-sdk options.resume, LangGraph checkpointer; leave raw Messages/Responses gap unless a documented Conversations-style API is wired. Do not invent APIs.

---

Implementations #

Status badges match ../GAPS.md: verified = tripwire PASS; linked — pending keys+PASS = path present, not done; gap / stub unchanged. Do not treat linked as verified.

Folder	Path	Status	Notes
claude/	—	gap	(raw Messages has no durable store)
openai/	—	gap	(raw Responses history is client-managed unless Conversations API — not wired here)
gemini/	—	gap
letta/	index.ts	verified	createSession → dispose → resumeSession(agentId)
eve/	—	gap	(durable sessions are a product pitch)
langchain/	index.py	linked — pending keys+PASS	SqliteSaver + thread_id (file-backed)
openai-agents/	index.ts	linked — pending keys+PASS	OpenAIConversationsSession + reconnect by conversationId
claude-agent-sdk/	index.ts	linked — pending keys+PASS	Capture session_id → options.resume
notion-agent/	index.ts	linked — pending keys+PASS	Continue with session_id (TS)
gemini-managed/	index.ts	linked — pending keys+PASS	previous_interaction_id + environment (TS)
claude-managed/	index.ts	linked — pending keys+PASS	Two turns on the same session id (TS)
cloudflare-agents/	index.ts → MemoryAgent in index.ts	linked — pending keys+PASS	this.setState / DO instance name