feat(hermes): add on_pre_compress hook to preserve context before compression

[pyrate-llama]

Summary

• Adds an on_pre_compress lifecycle hook to the Hermes integration that fires before context window compression
• Captures the last 10 user/assistant messages and persists them to the Hindsight knowledge graph via retain()
• Runs the retain call in a background thread so it never blocks the compression pipeline

Motivation

When Hermes compresses its context window to save tokens, the original messages are summarized and the full text is discarded. This means potentially valuable context is lost forever. By hooking into the compression lifecycle, we can preserve those messages in Hindsight's knowledge graph where they remain searchable via recall() and reflect().

This was inspired by ByteRover's on_pre_compress implementation, adapted to use Hindsight's sync retain() API.

Details

• Messages are truncated to 500 chars each to keep the retain payload reasonable
• Tagged with [Pre-compression context] prefix for easy identification
• Requires hermes-agent >= 0.7.0 which supports the on_pre_compress hook
• On older hermes-agent versions, the hook is simply never called (safe to register)

Test plan

• Tested locally with Hermes v0.7.0 and Hindsight in local/embedded mode
• Verified pre-compression messages appear in knowledge graph after context compression
• Confirmed background thread does not block the compression pipeline