Memory
Authoring contract. Attachments compose around autonomy; they are not a fifth mode.
dataplane.memory_profiles is policy on the route (conversation, working, loop, long_term, TTL, isolation). It is not a memory store. ADP serves the row; AR honors working and loop on the run pin (agent-runtime/app/core/memory.py).
| Field | Catalogue | This Runtime |
|---|---|---|
working=session | Scratch for this run | After each stage, writes { "notes": [...], "slots": { "<stage_id>": <json> } } to runtime.runs.working. /turns reloads both |
loop=checkpoint | Crash cursor | Writes {step, stage_id, result, goal, resume_index} to runtime.runs.checkpoint. Failed runs resume from resume_index via /turns with {} or { "resume": true } |
conversation=session | Next turn should see prior utterances | Catalogue-only. No transcript store. /turns reloads notes/slots, not user/assistant text |
long_term=retrieve_only | Facts for a later journey | Catalogue-only. Not retrieve, not prefetch |
AFD freeze (session_id → pin) is route stickiness, not conversation memory.
goal vs slots vs notes
| Channel | Role |
|---|---|
goal | Immutable ingress (job payload / chat utterance). Never updated mid-run. |
slots | Structured JSON per completed stage (slots[stage_id]). Used for HTTP schema merge, branch/gate, prefetch pack, child agent projection. |
notes | Append-only strings for LLM stages (_user_blob). Optional payload["notes"] on HTTP. |
HTTP merge: goal ∪ prior-slot keys that appear on the next capability input_schema — see data.
Not a Shared Memory box
conversation and long_term need a fifth store (not AFD / ADP / AR / ACR). This repo does not have that box. Do not write transcripts or long-term facts onto the run pin.
Retrieval is a sibling table. Prefetch writes working.slots.prefetch; it is not long_term. See retrieve.
Omit the profile on one-shot LLM/jobs (email_summarize, llm_pipeline). Author it when later stages or a later turn must keep working memory (working=session) or when an open loop should record a cursor (loop=checkpoint). fee_explain authors working + loop.
Components: architecture. Scenario matrix: dataflow scenarios. Verification: dataflow-plan.md.