hermes-lcm

What hermes-lcm does

Hermes Agent's built-in compressor prunes older tool results and asks an auxiliary model to summarize the older conversation once the prompt crosses a threshold, so the compacted turns are no longer verbatim in the model's context. hermes-lcm replaces that step. It persists raw messages in a plugin-local SQLite store, compacts older context into depth-aware summary nodes, and condenses those into a hierarchical DAG as they accumulate.

The active prompt is assembled from the system prompt, the highest-value summaries and a protected fresh tail. The agent gets recall tools such as lcm_grep, lcm_expand and lcm_describe to search and page back through compacted material. Optional controls cover ignored or read-only sessions, externalized large payloads and redaction of API keys and passwords. The design follows the LCM paper by Ehrlich and Blackman and is inspired by lossless-claw for OpenClaw.

Key features

• SQLite message store that preserves raw messages before compaction
• Depth-aware summary DAG that keeps source lineage
• Recall tools including lcm_grep, lcm_recall, lcm_expand and lcm_status
• Session controls to ignore noisy sessions or keep them read-only using glob patterns
• Optional named redaction of API keys, bearer tokens, passwords and private keys
• Diagnostics through lcm_doctor and optional /lcm slash commands

When to use it

• Long Hermes Agent sessions where compaction would otherwise drop exact earlier details
• Finding a specific earlier message or tool output after the context was compacted
• Importing history from OpenClaw's lossless-claw, which the README covers in an import section

Who it is for: Hermes Agent users who run long sessions and need exact detail to survive context compaction.

How it fits with Hermes Agent

Built for Hermes Agent as a plugin that replaces the built-in context compressor with its own context engine.

FAQ