hermes - 💡(How to fix) Fix [Feature]: Pre-compression task state preservation hook [1 participants]

Problem or Use Case

When Hermes Agent's context compression fires (triggered by context_engine: compressor), it destroys context without any pre-hook to preserve task state. This caused a catastrophic data loss: 249 conversation turns were wiped because the compression API call (Google Gemini, no key configured) failed silently.

The existing memory_hook.py has a precompact trigger designed precisely for this — it returns {"decision": "block"} to pause compression until memories are saved. However, nothing in the Hermes runtime actually invokes this hook before initiating compression.

The result: users lose entire conversation histories with no recovery mechanism, even when backup systems like Hindsight/MemPalace are running.

Proposed Solution

Two-part fix:

1. Wire precompact into the compression cycle

In context_compressor.py (or equivalent), add a hook call before call_llm for summarization:

# Before attempting compression summary
hook_result = self.trigger_hook("precompact", session_state)
if hook_result.decision == "block":
    # Wait for memories to be saved externally, then retry
    return {"status": "deferred", "reason": "pending_memory_save"}

This mirrors how memory_hook.py was already designed to work.

2. Add sessionRetention as a first-class config option

sessionRetention:
  enabled: true
  checkpoint_before_compress: true
  min_checkpoint_interval: 5  # exchanges

This gives users explicit control over session persistence behavior, separate from record_sessions.

Alternatives Considered

• Periodic checkpoint cron (workaround): Running session_checkpoint.py every N minutes reduces blast radius but doesn't guarantee preservation before compression fires.
• Changing compression provider: Switching from Google to another provider fixes the API key issue but doesn't solve the fundamental race condition — compression can still fail, and context is still destroyed without warning.
• Relying on Hindsight/MemPalace: These are excellent secondary memory systems, but they're async/backups, not part of the compression lifecycle. They can't block compression.

Feature Type

• Performance / reliability (data loss prevention)
• Configuration option (new sessionRetention config block)

Scope

• Medium — requires modifying compression engine lifecycle + adding config schema
• Likely 2-3 files: context_compressor.py, config schema, and documentation

Debug context: Confirmed on Hermes Agent v0.9.0 (2026.4.13). The precompact trigger in memory_hook.py exists but has no caller in the compression pipeline.