Memory Sidecar

What Memory Sidecar does

The sidecar follows a simple loop. It reads the agent's state and session data from an AGENT_HOME directory, archives new sessions into gbrain and a session search index, rebuilds governance and curated knowledge-note indexes, generates tiered recall context for the next agent turn, and runs health and acceptance checks so failures stay visible. The agent core is left untouched.

Recall is layered across hot, warm, cold and curated knowledge notes, so project playbooks and wiki pages can influence later answers. Version 3.5.2 adds synthetic and private memory evaluation registries, five quality metrics, governance policy metadata and dry-run-first gbrain edge planning. The memory-quality cron block is registered paused by default. Two Hermes-oriented maintenance scripts, memory_watermark.py and memory_snapshot_backup.py, sit in the repository but are not installed by default.

Key features

• Runs beside the agent without patching it, driven by an AGENT_HOME path
• Archives sessions into gbrain and a session search index
• Layered recall across hot, warm, cold and curated knowledge notes
• Machine-readable health artifacts and acceptance checks
• Memory evaluation registries with five quality metrics in v3.5.2

When to use it

• Giving a long-running Hermes agent durable recall across many sessions
• Letting project playbooks and wiki notes influence later answers
• Monitoring whether archival, recall or compaction has silently stopped

Who it is for: Technical users who already run a Hindsight, gbrain and PostgreSQL environment and want external long-term memory for Hermes or other agents.

How it fits with Hermes Agent

It works with Hermes alongside Claude Code, Codex and Cursor through AGENT_HOME, and ships two Hermes-oriented maintenance scripts that are not installed by default.

Requirements: A Hindsight, gbrain and PostgreSQL environment; installation is driven by an AGENT_HOME path.

FAQ