memory-augmented-reply

name	memory-augmented-reply
description	Before generating an AI reply, recall relevant memories from DeepMemory and search related code from FlashMemory, then inject both as augmented context. This is the memory-enhanced RAG pipeline — combining symbolic recall with vector search for richer, history-aware responses.
metadata	{"version":"1.0.0","category":"cognitive-memory","tags":["rag","memory-recall","context-augmentation","deepmemory","flashmemory"]}

Memory-Augmented Reply — Recall + RAG Context Injection

This skill enriches AI responses by combining two context sources before reply generation:

1. DeepMemory recall — past decisions, user preferences, session history
2. FlashMemory search — relevant code functions and modules

Together they provide the AI with both "what we've learned" (memory) and "what exists in the codebase" (retrieval).

When to Use This Skill

• Before every AI reply in a multi-turn conversation
• When context matters — the user references past decisions or preferences
• When code is involved — the user asks about implementation details
• When continuity matters — cross-session memory should inform the reply

Prerequisites

Both MCP servers must be running:

{
  "mcpServers": {
    "deepmemory": {
      "command": "python",
      "args": ["-m", "deepmemory.mcp_server"]
    },
    "flashmemory": {
      "command": "python",
      "args": ["-m", "flashmemory.mcp_server"]
    }
  }
}

Instructions

Step 1: Recall Memories

Query DeepMemory across multiple dimensions:

# Session-level memories
deepmemory_recall(object_id="session:{session_id}")

# User-level long-term preferences
deepmemory_recall(object_id="user:{user_id}")

# Topic-specific memories (extract from user message)
deepmemory_recall(object_id="module:{inferred_module}")
deepmemory_recall(object_id="concept:{inferred_concept}")

Filter the recalled memories:

• Prioritize CANONICAL and PROVISIONAL_CONSENSUS status memories
• Include LOCAL_HYPOTHESIS if relevant and recent
• Deprioritize or skip CONTESTED memories (flag them as uncertain)
• Ignore PRIVATE_HINT with confidence < 0.3

Step 2: Search Code (if applicable)

If the user's message involves code, architecture, or technical questions:

flashmemory_search(
  query="{user_message}",
  project_dir="{project_dir}",
  top_k=5,
  search_type="functions"
)

If the query is about project structure or module-level concerns:

flashmemory_search(
  query="{user_message}",
  project_dir="{project_dir}",
  top_k=3,
  search_type="modules"
)

Step 3: Build Augmented Context

Format the recalled information as a structured context block:

## Historical Decisions & Observations

- **[packaging_tool_preference]** (PROVISIONAL_CONSENSUS, confidence: 0.85)
  Claim: user prefers uv over pip — Source: user:alice, session-001

- **[lock_strategy_decision]** (LOCAL_HYPOTHESIS, confidence: 0.72)
  Claim: use RLock for recursive acquisition — Source: coder-agent

## Relevant Code

- **HandleUpload** (handlers package) — score: 0.89
  File: pkg/upload/handler.go:42
  Description: Handles multipart file upload with size validation

- **RetryWithBackoff** (utils package) — score: 0.76
  File: internal/utils/retry.go:15
  Description: Exponential backoff retry wrapper

Step 4: Inject into System Prompt

Prepend the augmented context to the system prompt before generating the reply. Do NOT generate the reply yourself — return only the context block for the dialog engine to use.

Memory Status Priority

Status	Priority	Usage
CANONICAL	Highest	Treat as established fact
PROVISIONAL_CONSENSUS	High	Reliable, multiple confirmations
LOCAL_HYPOTHESIS	Medium	Single-source, use with caveat
CONTESTED	Low	Flag as uncertain, present both sides
PRIVATE_HINT	Lowest	Only if nothing better available

Examples

Example: User Asks About Upload Feature

User message: "上传功能现在支持什么文件格式？"

Memory recall returns:

• module:upload:config → "支持 jpg/png/pdf，最大 10MB"（CANONICAL）
• user:preference:upload → "用户之前要求支持 webp 格式"（LOCAL_HYPOTHESIS）

Code search returns:

• ValidateFileType() in pkg/upload/validator.go — score 0.91
• HandleUpload() in pkg/upload/handler.go — score 0.85

Augmented context:

## Historical Decisions
- [upload_format_config] (CANONICAL): 支持 jpg/png/pdf，最大 10MB
- [upload_format_request] (LOCAL_HYPOTHESIS): 用户曾要求支持 webp

## Relevant Code
- ValidateFileType (upload.validator): 文件类型校验逻辑 — validator.go:28
- HandleUpload (upload.handler): 上传处理入口 — handler.go:42