C
ChaoBro

日本語版: OpenClaw v2026.4.29: Memory System Evolves from Retrieval-Based Recall to Person-Aware Wiki

日本語版: OpenClaw v2026.4.29: Memory System Evolves from Retrieval-Based Recall to Person-Aware Wiki

この記事は日本語版です。言語ルートを完全にするため、本文は既存の標準原稿をベースにしています。


Key Takeaways

OpenClaw released two updates within 48 hours (v2026.4.27 → v2026.4.29). The core of this update is a paradigm shift in the memory system — evolving from keyword-based passive retrieval to an agent-driven, actively built and maintained person-aware Wiki system.

This is a critical step for OpenClaw toward becoming a "long-term companion agent."

What Happened

Person-Aware Wiki

Previously, OpenClaw's memory system relied on vector retrieval: the agent searched the memory store for fragments most relevant to the current conversation. The limitations of this approach include:

  • Memories are fragmented, lacking structured connections
  • Unable to understand relationship networks between people
  • Memory sources are not traceable

The new Wiki system introduces three core mechanisms:

Mechanism Function Value
Person Card Agent automatically creates profiles for people mentioned in conversations Centralized management of person-related information
Relationship Graph Tracks connections between people Understands social context
Source Tracing Each memory is labeled with its source and evidence type Improves memory credibility

Source Tracing and Evidence Types

Every memory entry now carries metadata annotations:

memory: "User Zhang San prefers Python over JavaScript"
source: "2026-04-28 conversation"
evidence_type: "explicit"  # explicit | inferred | assumed
confidence: 0.92

Evidence types are categorized into three tiers:

  • explicit: User explicitly stated (high confidence)
  • inferred: Agent inferred from context (medium confidence)
  • assumed: Agent hypothesized based on patterns (needs verification)

Active Memory Enhancements

Active Memory (the working memory for the current conversation) gains two new capabilities:

  1. Conversation ID Filtering: Distinguishes memories across different conversation contexts, preventing context pollution
  2. Persistence Tagging: Marks specific memories for long-term retention, exempting them from automatic cleanup policies

Why It Matters

1. Memory Quality > Memory Quantity

Previous agent memory systems generally pursued "remembering more" — larger vector stores, longer context windows. OpenClaw's shift signals the industry starting to focus on "remembering more accurately":

Dimension Old Paradigm New Paradigm
Core Metric Recall Precision + Credibility
Storage Flat Vectors Structured Wiki + Graph
Update Mechanism Passive Append Active Build & Maintain
Expiration Policy Time Decay Evidence Type Driven

2. The Value of Relationship Graphs

In team collaboration scenarios, relationship graphs are particularly useful:

  • Project Management: The agent knows Zhang San is the backend lead and Li Si is the product manager, adjusting communication strategies accordingly
  • Customer Service: Identifies returning customers' historical interaction patterns to provide personalized service
  • Knowledge Management: Associates knowledge snippets with their creators for easier sourcing and updating

3. The Pace of Two Updates in Two Days

OpenClaw's consecutive updates within 48 hours (4.27 added computer-use capabilities, 4.29 upgraded the memory system) reflect an aggressive iteration speed for an open-source agent project. Behind this lies:

  • Community-driven rapid iteration
  • Modular architecture enabling independent module updates
  • Agent memory systems being the current competitive focal point

Comparison with Similar Projects

Project Memory Type Structured Person-Aware Source Tracing Open Source
OpenClaw Wiki + Graph
Claude Memory Retrieval-based
ChatGPT Memory Retrieval-based ⚠️ Partial
Hermes Agent Skill Management
Claude Managed Agents Session Memory

Recommendations

For OpenClaw Users:

  • Upgrade to v2026.4.29 to get the new memory system
  • Check existing memory library migration status — old vector memories may need re-indexing
  • Leverage evidence type filtering, prioritizing explicit memories for critical decisions

For Agent Developers:

  • The person-aware Wiki design pattern is worth借鉴, especially the source tracing mechanism
  • Consider introducing evidence type annotations in your own agents to improve decision transparency

For Technology Selection:

  • Need long-term memory + person management → OpenClaw currently leads
  • Only need short-term context → Built-in Claude/GPT memory is sufficient
  • Need skill management → Hermes Agent's Curator system is more mature