TencentDB Agent Memory solves a core problem for developers building long‑horizon AI agents: as agents run more steps, their context windows fill with verbose tool logs, search results and error traces, causing token bloat and unreliable recall. Traditional memory stacks flatten everything into a vector store, forcing a blind similarity search across disconnected fragments and losing the hierarchical structure that helps agents reason efficiently.
The system introduces a symbolic short‑term memory layer paired with a four‑tier semantic pyramid for long‑term storage. Verbose logs are offloaded to plain markdown files under refs/*.md while a compact Mermaid task canvas stays in the agent’s context window. When detail is needed, the agent looks up a node_id, greps the corresponding file and retrieves the raw text—providing deterministic drill‑down without wasting tokens.
Long‑term memory organizes information into L0 Conversation (raw dialogue), L1 Atom (atomic facts in JSONL), L2 Scenario (scene blocks in markdown) and L3 Persona (user profile in persona.md). The agent first queries the Persona layer for high‑level preferences, then drills down to Scenario, Atom or Conversation only when finer detail is required. This preserves evidence at the bottom while keeping structural guidance at the top.
Installation is straightforward. For OpenClaw users, add the npm package @tencentdb-agent-memory/memory-tencentdb, set the enabled flag in the openclaw.json config, and restart the gateway. The default backend uses local SQLite with the sqlite‑vec extension, requiring no external API. Hermes users can pull a pre‑built Docker image that bundles the agent, the memory plugin and the TDAI Memory Gateway, configuring model endpoints via environment variables.
Benchmarks show measurable gains: integrating the plugin with OpenClaw raises WideSearch pass rate from 33% to 50% while cutting token use by 61%; SWE‑bench improves from 58.4% to 64.2% with a 33% token reduction; AA‑CLR climbs from 44.0% to 47.5% and PersonaMem accuracy jumps from 48% to 76%. Retrieval defaults to a hybrid BM25 + vector strategy with Reciprocal Rank Fusion, falling back to pure keyword or embedding modes if desired, and times out gracefully after five seconds.
Developers get two runtime tools—tdai_memory_search and tdai_conversation_search—that return node_id and result_ref for traceback, enabling transparent debugging. All artifacts remain human‑readable under ~/.openclaw/memory‑tdai/, supporting white‑box inspection.
By combining symbolic compression, layered storage and easy local deployment, TencentDB Agent Memory gives teams a practical way to keep agents focused, reduce costs and improve reliability in long‑running tasks. #AI #Product #ProductManagement #UX #Innovation #Productivity #Technology #Startups #MachineLearning #OpenSource
