Community Implementation: Hermes ContextEngine plugin for TencentDB offload V2 API

[ysun0804]

Community ContextEngine Implementation

Following up on #202 (Hermes Context Engine support), we built and open-sourced a feature-complete Python ContextEngine plugin for Hermes Agent that uses TencentDB Gateway's offload V2 API.

Repo

https://github.com/ysun0804/tencentdb-offload

What it does

Aligns with the official OpenClaw offload implementation, providing:

• post_tool_call hook — async tool_pair ingest after each tool call (L1 data source)
• pre_llm_call hook — heartbeat filtering + MMD canvas injection + incremental compression
• L3 4-level compression — fastpath → mild → aggressive → emergency (via Gateway compact API)
• L2 Mermaid canvas injection — via Gateway query-mmd endpoint
• Reclaimer — stale session cleanup
• SessionRegistry + SessionState — multi-session state management with LRU eviction
• Graceful degradation — tail-truncation fallback when Gateway is unavailable
• Zero dependencies — pure Python stdlib (urllib, json, threading, logging)

Key implementation details

1. Hermes ContextEngine ABC doesn't have assemble() or afterToolCall — we use pre_llm_call and post_tool_call hooks as equivalents
2. context_window is dynamically calculated (tokens / 0.7) to make Gateway's ratio land in the mild compression range
3. ingest_before_compact preserves full messages before truncation

Tests

25/25 passing (14 core + 11 feature-specific)

Compatibility

• TencentDB Agent Memory v1.0.0+ (offload V2 API)
• Hermes Agent v0.17.0+
• Python 3.10+

Happy to collaborate if the team is interested in an official integration.

[Maxwell-Code07]

Thanks for your interest and contributions. We’ll review and get back to you shortly.