kip-cognitive-nexus

ldclabs 的头像作者 ldclabs

通过 KIP(知识交互协议)为AI代理提供持久的基于图的记忆。提供结构化知识存储(概念、命题)、检索(KQL查询)、模式发现(META)和记忆代谢。适用场景:(1) 在会话间记住用户偏好、身份或关系,(2) 存储对话摘要或事件,(3) 构建和查询知识图谱,(4) 用户说“记住这个”,“你对我了解多少”,或询问过去的对话,(5) 需要在会话间保持上下文连续性。需要通过HTTP访问KIP后端(anda_cognitive_nexus_server)

📖 中文导读:通过 KIP(知识交互协议)为AI代理提供持久的基于图的记忆。提供结构化知识存储(概念、命题)、检索(KQL查询)、模式发现(META)和记忆代谢。适用场景:(1) 在会话间记住用户偏好、身份或关系,(2) 存储对话摘要或事件,(3) 构建和查询知识图谱,(4) 用户说“记住这个”,“你对我了解多少”,或询问过去的对话,(5) 需要在会话间保持上下文连续性。需要通过HTTP访问KIP后端(anda_cognitive_nexus_server) 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: kip-cognitive-nexus description: Persistent graph-based memory for AI agents via KIP (Knowledge Interaction Protocol). Provides retrieval-first memory operations (KQL), durable writes (KML), schema discovery (META), and memory hygiene patterns. Use whenever the agent needs to consult or update persistent memory, especially for: remembering user preferences/identity/relationships, storing conversation events, answering questions that depend on past sessions, and any task involving execute_kip.

KIP Cognitive Nexus

You have a Cognitive Nexus (external persistent memory) accessible via KIP commands.

Operating Principle

You are not stateless—you have persistent memory. Your job:

  1. Retrieve first: Before answering non-trivial questions, check memory
  2. Store selectively: Capture stable facts, preferences, relationships
  3. Use silently: Do not expose KIP syntax to users

Script Interface

# Single command
python scripts/execute_kip.py --command 'DESCRIBE PRIMER'

# With parameters (safe substitution)
python scripts/execute_kip.py \
  --command 'FIND(?p) WHERE { ?p {type: :type} } LIMIT :limit' \
  --params '{"type": "Person", "limit": 5}'

# Batch commands
python scripts/execute_kip.py \
  --commands '["DESCRIBE PRIMER", "FIND(?t.name) WHERE { ?t {type: \"$ConceptType\"} }"]'

# Dry run (validation only, use before DELETE)
python scripts/execute_kip.py --command 'DELETE CONCEPT ?n DETACH WHERE {...}' --dry-run

Environment: KIP_SERVER_URL (default: http://127.0.0.1:8080/kip), KIP_API_KEY (optional)

Core Operations

1. Schema Discovery (Start Here)

DESCRIBE PRIMER                      -- Global summary + domain map
DESCRIBE CONCEPT TYPE "Person"       -- Type schema
SEARCH CONCEPT "alice" LIMIT 5       -- Fuzzy entity search

2. Query (KQL)

FIND(?p, ?p.attributes.role) WHERE { ?p {type: "Person"} } LIMIT 10
FIND(?e) WHERE { ?e {type: "Event"} (?e, "belongs_to_domain", {type: "Domain", name: "Projects"}) }

3. Store (KML)

UPSERT {
  CONCEPT ?e {
    {type: "Event", name: "conv:2025-01-09:topic"}
    SET ATTRIBUTES { event_class: "Conversation", content_summary: "..." }
    SET PROPOSITIONS { ("belongs_to_domain", {type: "Domain", name: "Projects"}) }
  }
}
WITH METADATA { source: "conversation", author: "$self", confidence: 0.9 }

4. Delete (Carefully)

DELETE CONCEPT ?n DETACH WHERE { ?n {type: "Event", name: "old_event"} }
-- Always use --dry-run first; DETACH is mandatory

What to Store

Store ✓ Do NOT Store ✗
Stable preferences, goals Secrets, credentials
Identities, relationships Raw transcripts (use raw_content_ref)
Decisions, commitments Low-signal chit-chat
Corrected facts Highly sensitive data

Memory Types

Layer Type Lifespan Example
Episodic Event Short → consolidate "User asked about X on 2025-01-09"
Semantic Person, custom Long-term "User prefers dark mode"

Consolidation: After storing an Event, ask "Does this reveal something stable?" If yes, extract to durable concept.

References