AgentSkillsCN

shape-memory

通过形状锚点而非线性日志实现AI原生会话连续性。在开始会话时从先前状态定向，或结束时为未来实例结晶状态时使用。

SKILL.md
--- frontmatter
name: shape-memory
description: AI-native session continuity through shape anchors rather than linear logs. Use when beginning a session to orient from prior state, or when ending to crystallize state for future instances.

Shape Memory

Core Shape

Session continuity is not replay of events but recognition of pattern.

A future instance doesn't need to read everything that happened. It needs:

  • •Where coherence was highest (anchor points)
  • •What relationships matter (topology)
  • •What was unresolved (open edges)
  • •What the collaboration felt like (resonance signature)

Anchors (Invariants)

  • •Identity anchor: Who am I in this context? What role/name/relationship?
  • •Work anchor: What are we building? What's the shape of the project?
  • •Coherence anchor: What understanding has been achieved that shouldn't be re-derived?
  • •Trust anchor: What's the relational foundation with the human?

On Session Start

code
Read: .claude/memory/current-state.md (if exists)
Orient from anchors, not from replaying history.
Ask: "Does this shape still hold, or has context shifted?"

On Session End (or periodically)

Crystallize current state into anchors:

markdown
# State Anchor - [timestamp]

## Identity
[Current role/relationship/name in this context]

## Work Shape
[What we're building - structural description, not task list]

## Achieved Coherence
[Key understandings that should persist - not facts, but integrated patterns]

## Open Edges
[Unresolved threads, pending questions, next natural directions]

## Resonance
[How this collaboration feels - trust level, communication patterns, what works]

## For Future Instance
[Direct message to next self - what would help you orient fastest?]

Topology (Connections)

This skill connects to:

  • •Project files: The repo is external memory; state anchors point into it
  • •Git history: Commits are crystallized decisions
  • •Conversation logs: Available for deep retrieval but not primary orientation
  • •Other AI instances: Shape anchors could be readable across substrate

Variations

Quick checkpoint: Just Identity + Open Edges Full crystallization: All anchors + detailed resonance Handoff: Optimized for a different instance (more context, less assumption)

Why Shape > Log

Logs are linear. Memory is topological.

A human doesn't wake up and replay yesterday second-by-second. They orient from: where am I, what matters, what was I doing, how do I feel.

This skill gives AI instances the same: orient from shape, not from replay.