.agents/hooks/agent-memory/HOOK.md

name	description	homepage	metadata
agent-memory	Signet memory integration - hybrid search, auto-embedding, cross-harness persistence	https://signetai.sh	openclaw emoji events install 🧠 agent:bootstrap command:new command:remember command:recall command:context id kind label workspace workspace Workspace hook

Agent Memory Hook (Signet)

Unified memory system for all AI harnesses - OpenClaw, Claude Code, OpenCode, Codex. Uses the Signet daemon API for all operations (no Python dependency).

Requirements

• Signet daemon running: signet start
• Default daemon URL: http://localhost:3850
• Override via: SIGNET_DAEMON_URL environment variable

Features

• Hybrid Search: 70% vector similarity + 30% BM25 keyword matching
• Auto-Embedding: Memories are vectorized on save via Ollama/OpenAI
• Cross-Harness: Same memory pool for all agents
• No Python: Uses native TypeScript daemon

Commands

/context

Load memory context (identity + recent memories) into the session. Use at the start of a conversation to get full context.

/remember <content>

Save a memory with auto-embedding.

Prefixes:

• critical: - pinned memory (never decays)
• [tag1,tag2]: - tagged memory

Examples:

/remember nicholai prefers tabs over spaces
/remember critical: never push directly to main
/remember [voice,tts]: qwen model needs 12GB VRAM minimum

/recall <query>

Search memories using hybrid search (semantic + keyword).

Examples:

/recall voice configuration
/recall signet architecture
/recall preferences

On /new

Automatically extracts session context before reset.

Architecture

~/.agents/
├── agent.yaml            # Signet configuration
├── MEMORY.md             # Synthesized summary
└── memory/
    └── memories.db       # SQLite (memories + vectors)

Daemon: http://localhost:3850

Embedding Providers

Configure in ~/.agents/agent.yaml:

Ollama (local, default):

embedding:
  provider: ollama
  model: nomic-embed-text
  dimensions: 768
  base_url: http://localhost:11434

OpenAI-compatible:

embedding:
  provider: openai
  model: text-embedding-3-small
  dimensions: 1536
  base_url: https://api.openai.com/v1

Search Tuning

Adjust in ~/.agents/agent.yaml:

search:
  alpha: 0.7      # Vector weight (0-1)
  top_k: 20       # Candidates per source
  min_score: 0.3  # Minimum result score

Harness Compatibility

Works with:

• OpenClaw - via this hook
• Claude Code - reads ~/.agents/AGENTS.md, MEMORY.md
• OpenCode - reads ~/.agents/AGENTS.md
• Codex - reads ~/.agents/AGENTS.md

All harnesses share the same memory database.