From dd73711b689ea12b6d08d9d20440a5ab4a586479 Mon Sep 17 00:00:00 2001
From: Shinwoo PARK <shinwoo.park@psw.kr>
Date: Mon, 11 May 2026 23:14:00 +0900
Subject: [PATCH] docs: document conversation memory pipeline

---
 API-Reference.md | 8 ++++++++
 Home.md          | 1 +
 2 files changed, 9 insertions(+)

diff --git a/API-Reference.md b/API-Reference.md
index ec6bd46..dfd15aa 100644
--- a/API-Reference.md
+++ b/API-Reference.md
@@ -140,6 +140,7 @@ Class-based entrypoint for inbound replies, proactive openings, and entry listin
 ### `replyToConversation(db, input)`
 
 Persists an inbound message and generates a DM-style response turn.
+When `input.memoryPipeline` is provided, the turn can also classify durable inbound/outbound messages and persist extracted IdentityDB facts back into the persona space.
 
 ```ts
 replyToConversation(db: IdentityDB, input: ReplyToConversationInput): Promise<ConversationTurnResult>
@@ -148,6 +149,7 @@ replyToConversation(db: IdentityDB, input: ReplyToConversationInput): Promise<Co
 ### `startConversation(db, input)`
 
 Generates a proactive outbound opening turn with no inbound user message.
+The same optional `memoryPipeline` can classify and extract durable proactive outbound memories.
 
 ```ts
 startConversation(db: IdentityDB, input: StartConversationInput): Promise<ConversationTurnResult>
@@ -167,6 +169,12 @@ Required model roles for turn generation:
 - `contextualMemoryModel: StructuredModelAdapter`
 - `responseModel: StructuredModelAdapter`
 
+Optional long-term memory pipeline:
+
+- `memoryPipeline.classifierModel: StructuredModelAdapter`
+- `memoryPipeline.extractorModel: StructuredModelAdapter`
+- `memoryPipeline.source?: string`
+
 ## `memory`
 
 ### `FactDraftMemoryStore`
diff --git a/Home.md b/Home.md
index 7ff4eee..63c0eb1 100644
--- a/Home.md
+++ b/Home.md
@@ -17,6 +17,7 @@ It is not a finished chatbot product. It is a **framework-first runtime library*
 - availability persistence with schedule/manual/tool overrides
 - availability snapshots with current and next-transition resolution
 - DM-style conversation orchestration for inbound replies and proactive openings
+- optional two-stage memory extraction for durable inbound/outbound conversation facts
 - human-like first-reply and typing delay helpers
 
 ## What is still planned