TencentTencentDB Agent Memory is a memory plugin for OpenClaw and Hermes agents that runs entirely on the local machine, defaulting to a SQLite + sqlite-vec backend. It addresses two problems: verbose tool logs that flood the context window during long tasks, and the need to re-explain workflows, project background, and conventions to an agent across sessions.
For long-term memory it builds a four-tier semantic pyramid instead of a flat vector store: L0 raw conversations, L1 atomic facts, L2 scenario blocks, and L3 user personas. Recall starts at the persona or scenario level and drills down to atoms or raw dialogue only when specific facts matter. Lower tiers are persisted in the database for full-text retrieval, while upper tiers are written as Markdown files that a person can inspect.
For short-term memory it offloads full tool outputs to external Markdown files, extracts step-level summaries to JSONL, and keeps only a compact Mermaid task canvas in context. Each canvas node carries a node_id, so the agent can retrieve the underlying raw text on demand rather than relying on lossy summaries.
Reported results against OpenClaw on continuous long-horizon sessions:
| Memory | Benchmark | Baseline success | With plugin | Baseline tokens | With plugin tokens |
|---|---|---|---|---|---|
| Short-term | WideSearch | 33% | 50% | 221.31M | 85.64M |
| Short-term | SWE-bench | 58.4% | 64.2% | 3474.1M | 2375.4M |
| Short-term | AA-LCR | 44.0% | 47.5% | 112.0M | 77.3M |
| Long-term | PersonaMem | 48% | 76% | n/a | n/a |
Features
- Layered long-term memory: L0 conversation → L1 atom → L2 scenario → L3 persona, with drill-down from abstraction to evidence
- Symbolic short-term memory: tool logs offloaded to files and replaced in context by a Mermaid canvas with node_id references
- Hybrid recall: keyword, embedding, or RRF-fused hybrid strategy with configurable result counts and character budgets
- Automatic pipeline: conversation capture, extraction every N turns, idle-timeout triggers, scenario aggregation, and persona generation every N memories
- Context offload thresholds: mild and aggressive compression ratios tied to the context window size
- Local by default: SQLite + sqlite-vec storage with no external API dependency, plus optional remote OpenAI-compatible embeddings (including sendDimensions toggle for models such as BGE-M3)
- Standalone LLM mode: run the L1/L2/L3 passes with a designated API instead of the host agent's model
- Hermes gateway: an HTTP sidecar exposing capture, search, recall, and health endpoints, shipped as a Docker image or plugged into an existing install, including Windows native
- Gateway security: opt-in Bearer API key and CORS allow-list, with GET /health left open for orchestrator probes
- Retention and exclusion controls: per-agent glob exclusions and L0/L1 retention days
- Tokenizer options: BM25 tokenizer for Chinese (jieba) or English
