symbolism-extract

name	symbolism-extract
description	Extract symbolic order from conversations: identify key concepts as Signs, anchor them to ObjectRefs, and record interpretant claims via DeepMemory MCP. This is the "semiotics deconstructs semantics" engine — it turns natural language into structured, contestable memory units instead of opaque vectors.
metadata	{"version":"1.0.0","category":"cognitive-memory","tags":["semiotics","symbol-extraction","deepmemory","memory-ingest"]}

Symbolism Extract — Semiotic Deconstruction Engine

This skill automatically extracts symbolic structures from dialog messages and persists them as DeepMemory MemoryUnits. It implements the core idea of "using semiotics to deconstruct semantics" — every observation becomes a structured triple of Sign + ObjectRef + Interpretant, not just a vector.

When to Use This Skill

• After each AI reply — extract key decisions, observations, preferences
• After user messages — capture user intent, corrections, stated preferences
• After code changes — record the reasoning behind architectural decisions
• When switching context — summarize the current session's key insights

Prerequisites

The DeepMemory MCP Server must be running and registered:

{
  "mcpServers": {
    "deepmemory": {
      "command": "python",
      "args": ["-m", "deepmemory.mcp_server"],
      "env": {"DEEPMEMORY_STORE": "./artifacts/memory.jsonl"}
    }
  }
}

Instructions

Step 1: Analyze the Message

Read the input message and identify extractable triples:

1. Signs — Key concepts, preferences, decisions, or observations

   • Examples: "packaging tool preference", "auth middleware pattern", "retry strategy"
   • Use snake_case English for canonical_sign
   • List natural-language variants in the user's language

2. ObjectRefs — What concrete object does this sign point to?

   • object_id: hierarchical identifier (e.g. user:preference:packaging, module:auth:middleware)
   • object_type: one of concept, module, user_preference, decision, bug, feature
   • anchor_kind: symbol, locator, or config
   • anchor_locator: file path or logical location

3. Interpretants — Who believes what, and how strongly?

   • actor_ref: the speaker (user:alice, coder-agent, reviewer-agent)
   • operational_claim: one-sentence claim in the speaker's language
   • confidence: inferred from language cues
     • Definitive statements ("always use", "never do") → 0.85-0.95
     • Preferences ("I prefer", "I like") → 0.70-0.85
     • Tentative ("maybe", "could try") → 0.40-0.60
     • Questions or uncertainty → 0.20-0.40

Step 2: Check for Existing Memories

Before ingesting, call deepmemory_recall for each object_id to check if this concept already has memories:

deepmemory_recall(object_id="user:preference:packaging")

• No existing memories → proceed to ingest
• Existing + consistent → skip (don't duplicate)
• Existing + supplementary → ingest as new observation
• Existing + contradictory → revise the old memory via deepmemory_revise, or ingest as a new competing claim (the game-theory-debate skill will resolve it)

Step 3: Ingest New Memories

For each extracted triple, call the DeepMemory MCP tool:

deepmemory_ingest(
  summary="user prefers uv over pip for dependency management",
  object_id="user:preference:packaging",
  object_type="user_preference",
  anchor_kind="config",
  anchor_locator="pyproject.toml",
  template="task",
  actor_ref="user:alice",
  source_ref="session-001-turn-005",
  context='{"session_id": "abc", "turn": 5}'
)

Step 4: Report Extraction Results

Output a structured summary:

{
  "extracted": [
    {
      "sign": "packaging_tool_preference",
      "object_id": "user:preference:packaging",
      "claim": "用户偏好uv而非pip",
      "confidence": 0.85,
      "action": "ingested"
    }
  ],
  "skipped": 0,
  "revised": 0
}

Template Selection Guide

Scenario	Template	Source Kind
User states a preference or observation	task	task_note
Architectural decision or design choice	architecture	architecture_decision
Bug fix reasoning or incident response	repair	incident_repair

Examples

Example 1: User Preference

Input: "我更喜欢用 uv 而不是 pip 来管理 Python 依赖"

Extraction:

• Sign: packaging_tool_preference
• ObjectRef: user:preference:packaging (type: user_preference)
• Interpretant: actor=user:alice, claim="偏好uv管理Python依赖", confidence=0.82

Example 2: Architectural Decision

Input: "We should use RLock instead of Lock here to allow recursive acquisition"

Extraction:

• Sign: lock_strategy_decision
• ObjectRef: module:cache:concurrency (type: decision)
• Interpretant: actor=coder-agent, claim="use RLock for recursive lock acquisition", confidence=0.88

Example 3: Code Change Reasoning

Input: "Narrowed the lock scope to reduce contention under high concurrency"

Extraction:

• Sign: lock_scope_optimization
• ObjectRef: module:upload:service (type: decision)
• Interpretant: actor=coder-agent, claim="narrow lock scope to reduce contention", confidence=0.90