Hermes-LCM is a context-engine plugin for Hermes Agent that replaces its built-in one-shot compression with Lossless Context Management (LCM), based on the LCM paper by Ehrlich and Blackman. The built-in compressor summarizes older conversation and rebuilds the prompt from that summary plus a recent tail, so compacted turns leave the active context. Hermes-LCM keeps the live prompt bounded but preserves every raw message and gives the agent tools to drill back into compacted material.
Messages are persisted in a plugin-local SQLite store with full-text search metadata. Older context is compacted into depth-aware summary nodes, and those summaries are condensed into a hierarchical directed acyclic graph (DAG) as they accumulate. The active prompt is assembled from the system prompt, the highest-value summaries, and a protected fresh tail. Summaries keep source lineage, and oversized tool or media payloads are externalized behind stable references that can be expanded later.
Three opt-in feature families extend it from compression into a memory system: large-output externalization with context-budget controls, temporal memory with day, week, and month rollups and natural-time recall, and semantic retrieval that adds embedding-backed and hybrid search using free-tier cloud or fully local embedding providers. Further opt-in tools store structured assertions (facts, preferences, commitments) with exact message citations and compute dates, counts, sums, and differences over cited evidence. The plugin also bundles a recall skill and injects a recall policy into the user-message context, and an import script backfills history from OpenClaw and lossless-claw databases.
Features
- Summary DAG: older turns compact into depth-aware summary nodes that condense hierarchically while keeping source lineage
- Raw message store: every message is preserved in SQLite with full-text search before compaction
- Recall tools: 15 agent tools to search, recall across conversations, expand summaries, load transcripts, and inspect the DAG
- Bounded recovery: raw messages, child summaries, and externalized payloads are paged rather than dumped into the prompt
- Cross-conversation recall: fuses full-text, summary-vector, and chunk-vector search with a recency prior, and works full-text-only without embeddings
- Temporal memory: day, week, and month rollups with retrieval by natural time period
- Evidence and computation: opt-in assertion storage and dependency-free computation over exact cited messages, failing closed on ambiguous input
- Payload and secret controls: externalizes oversized tool and media output and can redact API keys, tokens, passwords, and private keys before storage
- Session controls: ignore noisy sessions or keep them read-only with glob patterns
- Diagnostics: health and database checks plus optional operator commands with backup-first repair, cleanup, and in-place rotate
- OpenClaw import: backfills history from OpenClaw or lossless-claw databases and session transcripts with explicit provenance
