Rewind Memory — Persistent Bio-Inspired Memory for AI Agents

Persistent, bio-inspired memory for AI agents. 5-layer architecture (L0-L4) with BM25 keyword search, knowledge graph, vector similarity, and HybridRAG fusio...

Security Scan

The skill's code and runtime instructions match its stated purpose (a persistent, layered memory system); behaviour is coherent but it persistently captures workspace data and — if you opt into Pro — can send queued text to an external batch endpoint, so review config before enabling.

Purpose & Capability

Name/description align with the provided files and hooks. The package implements layered memory (search, KG, vectors), provides CLI integration, and registers hooks to capture and index session data — all consistent with a 'persistent memory' skill.

Instruction Scope

Runtime hooks capture file edits, command outputs, and user prompts and will index or store those as memory. UserPrompt injects prior context into system messages. This scope is consistent with the memory use-case but has high privacy impact (it may capture command outputs and snippets of files automatically). The setup flow can pull local models (ollama) and writes ~/.rewind/config.yaml.

Install Mechanism

No installer spec is embedded in the skill bundle. The SKILL.md instructs users to 'pip install rewind-memory' and to optionally pull models with Ollama; code files do not download arbitrary archives themselves. No suspicious external download URLs or shorteners are embedded in the included files.

Credentials

The skill does not require platform credentials but reads several environment/config values (CLAUDE_PLUGIN_ROOT, REWIND_DATA_DIR, REWIND_API_URL) and the user-writable ~/.rewind/config.yaml. Pro features rely on modal.extract_batch_url and modal.auth_token from that config to POST queued texts to an external Modal endpoint — this is justified by the Pro workflow but is sensitive and should only be enabled with trusted endpoints and tokens.

Persistence & Privilege

The skill registers hooks (PostToolUse, Stop, UserPromptSubmit) that run local Python scripts and write persistent data under ~/.rewind (and a queue directory). always:false (not force-enabled). The behavior is expected for a memory plugin, but it will persistently store session content and can autonomously inject memory into prompts.

Assessment

This skill appears to do what it says: it automatically captures edits, tool outputs, and conversation turns and stores them in a local ~/.rewind data directory; it will also search that memory and inject relevant context into prompts. Things to consider before enabling: - Privacy: Hooks may capture command outputs and file contents (including secrets accidentally printed to stdout). Audit what your tools output and avoid storing sensitive data. - Pro / cloud features: The Pro path sends queued text to an external Modal endpoint using an auth token stored in ~/.rewind/config.yaml; only configure this if you trust the remote service and keep the auth token secret. - Local binaries: The hooks call an external 'rewind' CLI and may call 'ollama pull' during setup; ensure you inspect/approve those tools and their network activity before running them. - Config review: Inspect ~/.rewind/config.yaml (or REWIND_DATA_DIR) before use to confirm providers, endpoints, and tokens. - Minimizing risk: If you want memory but not automatic capture, disable or remove the hooks, or decline to enable Pro/cloud features. If you have low tolerance for automated context injection, do not enable the UserPromptSubmit/PostToolUse hooks. Overall the package is internally coherent (benign), but it has meaningful privacy and exfiltration implications if you enable Pro or provide external endpoints — review configs and opt-ins carefully.

Rewind Memory — Persistent Memory for AI Agents

Give your agent memory that survives sessions, compactions, and restarts.

What It Does

• L0 Sensory Buffer — BM25 keyword search over all indexed content
• L1 System Memory — core files loaded every session
• L2 Orchestrator — fuses all layers with unified multi-signal scoring
• L3 Knowledge Graph — entity extraction, relationships, spreading activation
• L4 Vector Search — semantic similarity via sqlite-vec

Install

pip install rewind-memory

Quick Start

rewind doctor          # auto-diagnose + build index
rewind ingest ./docs/  # index your files
rewind search "what did we decide about auth?"
rewind watch-sessions  # real-time conversation capture

OpenClaw Integration

# Native hook (recommended — survives updates)
rewind-openclaw hook

# Or configure memory_search routing
rewind-openclaw setup

MCP Server

{
  "mcpServers": {
    "rewind": { "command": "rewind-mcp" }
  }
}

6 tools: memory_search, memory_store, memory_extract, memory_stats, memory_feedback, graph_traverse.

Pro Upgrade

Pro adds L5 (communications), L6 (documents), NV-Embed-v2 4096-dim cloud embeddings, cross-encoder reranking, and bio-inspired lifecycle (decay, consolidation, pruning).

pip install rewind-memory-pro

Links

• Documentation
• GitHub
• PyPI

Built by SARAI Defence. Patent pending.