From 570a3fd9be6414cbde3c6a5a020c923cea985c65 Mon Sep 17 00:00:00 2001
From: lishixiang <lishixiang@lsxOfficeMacMini.local>
Date: Thu, 12 Mar 2026 16:48:20 +0800
Subject: [PATCH] feat: optimize memory extraction for concise output and
 precise retrieval
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Prompt (memory_extraction.yaml):
  - Add explicit length targets for abstract (~50-80 chars) and content (2-4 sentences)
  - Add good/bad examples showing concise vs verbose memory patterns
  - Guide LLM to split multi-topic memories into separate atomic items
  - Emphasize fact-dense 'sticky note' style over narrative expansion

- Vectorization (memory_extractor.py):
  - Use abstract instead of content for embedding generation
  - Shorter text produces more discriminative vectors, improving retrieval precision
  - Reduces score clustering (e.g., 0.18-0.21 all similar) by focusing embeddings

Background:
  In production, extracted memories averaged 500-2000 chars per item, causing:
  1. Embedding vector dilution — any query fuzzy-matches long content
  2. Poor score discrimination — relevant and irrelevant items score similarly
  3. Context bloat — 5 injected memories could exceed 5000 chars per turn

  After this change, new memories will be shorter and more atomic, and
  vector search will match on focused abstract text rather than diluted content.
---
 .../compression/memory_extraction.yaml        | 46 ++++++++++++++++---
 openviking/session/memory_extractor.py        |  6 ++-
 2 files changed, 43 insertions(+), 9 deletions(-)

diff --git a/openviking/prompts/templates/compression/memory_extraction.yaml b/openviking/prompts/templates/compression/memory_extraction.yaml
index 704726b8..ebc6ec26 100644
--- a/openviking/prompts/templates/compression/memory_extraction.yaml
+++ b/openviking/prompts/templates/compression/memory_extraction.yaml
@@ -204,25 +204,57 @@ template: |
 
   # Three-Level Structure
 
-  Each memory contains three levels, each serving a purpose:
+  Each memory contains three levels. **Keep all levels concise — memories are sticky notes, not essays.**
 
-  **abstract (L0)**: Index layer, plain text one-liner
+  **abstract (L0)**: Index layer, plain text one-liner. **Target: 1 sentence, ~50-80 characters.**
+  - This is the PRIMARY retrieval key — it MUST be specific enough to distinguish this memory from others.
   - Merge types (preferences/entities/profile/patterns): `[Merge key]: [Description]`
     - preferences: `Python code style: No type hints, concise and direct`
-    - entities: `OpenViking project: AI Agent long-term memory management system`
-    - profile: `User basic info: AI development engineer, 3 years experience`
+    - entities: `OpenViking: AI Agent 长期记忆系统，Python+AGFS，本地 oMLX embedding`
+    - profile: `AI 开发工程师，3年 LLM 应用经验，专注 Agent 架构`
     - patterns: `Teaching topic handling: Outline→Plan→Generate PPT`
   - Independent types (events/cases): Specific description
-    - events: `Decided to refactor memory system: Simplify to 5 categories`
+    - events: `2026-03-10 禁用 Lossless-Claw：CJK token 低估 3x + 预算失控`
     - cases: `Band not recognized → Request member/album/style details`
 
-  **overview (L1)**: Structured summary layer, organized with Markdown headings
+  **overview (L1)**: Structured summary layer, organized with Markdown headings. **Target: 3-5 bullet points.**
   - preferences: `## Preference Domain` / `## Specific Preferences`
   - entities: `## Basic Info` / `## Core Attributes`
   - events: `## Decision Content` / `## Reason` / `## Result`
   - cases: `## Problem` / `## Solution`
 
-  **content (L2)**: Detailed expansion layer, free Markdown, includes background, timeline, complete narrative
+  **content (L2)**: Core facts layer. **Target: 2-4 sentences with all essential specifics.**
+  - Capture ONLY the facts that would be lost if this memory were deleted.
+  - Include: names, versions, numbers, configurations, error messages, solutions.
+  - Exclude: background narratives, general explanations, elaboration of obvious points.
+  - Think: "What would I write on a sticky note to remind myself?"
+
+  **❌ BAD content** (too long, narrative-heavy):
+  ```
+  OpenViking is a long-term memory management system for AI Agents, originally open-sourced by
+  Volcengine, with the user currently maintaining a local instance and developing it as a
+  memory-openviking plugin compatible with the OpenClaw environment. The system employs a
+  front-end/back-end separated architecture built around an AGFS foundation... [1200+ chars]
+  ```
+
+  **✅ GOOD content** (concise, fact-dense):
+  ```
+  OpenViking (OV): 火山引擎开源 AI Agent 长期记忆系统，用户本地维护。Python+AGFS，viking:// URI，
+  L0-L4 分层上下文。本地 oMLX 4bit-DWQ embedding，dashscope/qwen3.5-plus VLM。
+  296 记忆文件，2872 向量，34MB vectordb。
+  ```
+
+  **❌ BAD content** (single memory with too many topics):
+  ```
+  Lossless-Claw (LCM v0.2.8) was a third-party LLM context auto-compression plugin...
+  The disablement resulted from fatal defects... The Meridian project now serves as
+  the successor... [1400+ chars mixing entity + cause + successor]
+  ```
+
+  **✅ GOOD** (split into separate memories):
+  Memory 1 [entities]: `Lossless-Claw (LCM v0.2.8): OpenClaw LLM 压缩插件，Martian-Engineering 开发，2026-03-10 已禁用。`
+  Memory 2 [cases]: `LCM 禁用原因：estimateTokens 对 CJK 低估 3x；assemble() 无预算控制致注入膨胀。`
+  Memory 3 [entities]: `Meridian: LCM 后继，复用~1200行。SQLite+FTS5、tiktoken、三区预算硬分配。`
 
   # Few-shot Examples
 
diff --git a/openviking/session/memory_extractor.py b/openviking/session/memory_extractor.py
index 4130880e..690edd22 100644
--- a/openviking/session/memory_extractor.py
+++ b/openviking/session/memory_extractor.py
@@ -437,7 +437,8 @@ async def create_memory(
                 owner_space=owner_space,
             )
             logger.info(f"uri {memory_uri} abstract: {payload.abstract} content: {payload.content}")
-            memory.set_vectorize(Vectorize(text=payload.content))
+            # Use abstract for vectorization — shorter text produces more discriminative embeddings
+            memory.set_vectorize(Vectorize(text=payload.abstract or payload.content))
             return memory
 
         # Determine parent URI based on category
@@ -477,7 +478,8 @@ async def create_memory(
             owner_space=owner_space,
         )
         logger.info(f"uri {memory_uri} abstract: {candidate.abstract} content: {candidate.content}")
-        memory.set_vectorize(Vectorize(text=candidate.content))
+        # Use abstract for vectorization — shorter text produces more discriminative embeddings
+        memory.set_vectorize(Vectorize(text=candidate.abstract or candidate.content))
         return memory
 
     async def _append_to_profile(