Hybrid Memory

**Quick judgement** Hybrid Memory is a strong, practical choice for OpenClaw agents that need both semantic document recall and time‑aware context over long-running interactions. It’s especially useful if you already use or are willing to deploy Graphiti and want agents to reliably decide *which* memory system to query for a given question. **What it does well** - Combines OpenClaw’s built-in vector memory (`memory_search` + `memory_get`) with a temporal knowledge graph (Graphiti). - Provides a clear decision rule: temporal questions → Graphiti; document/content questions → memory_search; uncertain → try both and merge. - Supports temporal and entity‑tracking queries that plain embedding search is weak at ("when did we set up Slack?", "what projects involve Alice?"). - Offers concrete tooling patterns and an `AGENTS.md` template, making it easier to wire into existing agents. **Key risks / limitations** - **Operational complexity**: Requires Graphiti deployed via Docker and associated scripts (`graphiti-search.sh`, `graphiti-log.sh`), which adds infra overhead and potential points of failure. - **Data consistency**: Graphiti is only as good as what gets logged; if agents or daemons don’t consistently record facts, temporal answers will be incomplete or misleading. - **Privacy & governance**: Encourages logging potentially sensitive user facts into a long-lived graph; there’s no explicit access control, redaction, or retention strategy described. - **Environment assumptions**: Assumes a Unix-like environment with shell access and the Graphiti stack; not ideal for constrained or serverless runtimes. **Recommended scenarios** - Long-running assistants (project managers, team bots, research companions) where remembering *when* something happened or *how it changed over time* is important. - Agents that must search both markdown-based knowledge (e.g., `GOALS.md`, `memory/**/*.md`) and historical interactions or events. - Multi-user or multi-project environments where you want structured tracking by group ID (e.g., `main-agent`, `user-personal`). - Teams comfortable managing Docker stacks and external services, who want a principled hybrid memory layer rather than ad-hoc notes + embeddings alone.