Tencent has released TencentDB Agent Memory, an open-source memory system for AI agents. The project ships under the MIT license. It targets a problem familiar to anyone shipping long-horizon agents: context bloat and recall failure.
It is symbolic short-term memory along with layered long-term memory. It integrates with OpenClaw as a plugin and with the Hermes Agent through a Gateway adapter. The default backend is local SQLite with the sqlite-vec extension, so no external API is required.
Why agent memory is hard
Most current memory stacks shred data into fragments and dump them into a flat vector store. Recall then becomes a blind similarity search across disconnected fragments, with no macro-level guidance. The architecture rests on two pillars: memory layering and symbolic memory.
A 4-tier semantic pyramid
For long-term personalization, TencentDB Agent Memory builds a four-level pyramid instead of a flat log. The layers are L0 Conversation, L1 Atom, L2 Scenario, and L3 Persona. These correspond to raw dialogue, atomic facts, scene blocks, and a user profile.
The Persona layer carries day-to-day user preferences and is queried first. The system drills down to Atoms or raw Conversations only when finer detail is needed. Lower layers preserve evidence; upper layers preserve structure.
Storage is heterogeneous. Facts, logs, and traces are persisted in databases for full-text retrieval. Personas, scenes, and canvases are stored as human-readable Markdown files. Layered memory artifacts live under ~/.openclaw/memory-tdai/.
Symbolic short-term memory via Mermaid
Long-running agent tasks consume tokens through verbose tool logs, search results, code, and error traces. TencentDB Agent Memory addresses this through context offloading combined with symbolic memory.
Full tool logs are offloaded to external files under refs/*.md. State transitions are encoded in Mermaid syntax inside a lightweight task canvas. The agent reasons over the symbol graph in its context window.
When it needs the raw text, it greps for a node_id and retrieves the corresponding file. The Tencent dev team describes this as a deterministic drill-down from top-layer symbol to mid-layer index to bottom-layer raw text.
Benchmark numbers
Results are measured over continuous long-horizon sessions, not isolated turns. SWE-bench, for example, runs 50 consecutive tasks per session to simulate context-accumulation pressure.
On WideSearch, integrating the plugin with OpenClaw raises pass rate from 33% to 50%, a 51.52% relative improvement. Token usage drops from 221.31M to 85.64M, a 61.38% reduction.
On SWE-bench, success climbs from 58.4% to 64.2% while tokens fall from 3474.1M to 2375.4M, a 33.09% reduction. On AA-LCR, the success rate moves from 44.0% to 47.5%. Tokens drop from 112.0M to 77.3M, a 30.98% reduction.
For long-term memory, PersonaMem accuracy rises from 48% to 76%. Note: these numbers come from Tencent’s own evaluations.
Recall and retrieval
Retrieval defaults to a hybrid strategy. The system combines BM25 keyword search with vector embeddings, fused using Reciprocal Rank Fusion (RRF). Developers can switch to pure keyword or embedding mode through a config field. The BM25 tokenizer supports both Chinese (jieba) and English.
Default settings trigger an L1 memory extraction every five turns. A user persona is generated every 50 new memories. Recall returns five items by default with a 5-second timeout. On timeout, the system skips injection rather than blocking the conversation.
Installation and developer surface
The OpenClaw integration ships as a single npm package: @tencentdb-agent-memory/memory-tencentdb. The project requires Node.js 22.16 or higher. Enabling it takes one config flag. The plugin then handles conversation capture, memory extraction, scene aggregation, persona generation, and recall.
For Hermes, a Docker image bundles the agent, the plugin, and the TDAI Memory Gateway. The default model is Tencent Cloud’s DeepSeek-V3.2. Any OpenAI-compatible endpoint works through the MODEL_PROVIDER=custom flag.
Two tools are exposed to agents during a session: tdai_memory_search and tdai_conversation_search. Both return references with node_id and result_ref fields for traceback. A Tencent Cloud Vector Database (TCVDB) backend is also available as an alternative to local SQLite.
Marktechpost’s Visual Explainer
TencentDB Agent Memory — Preview
body{margin:0;padding:24px;background:#f5f5f5;font-family:-apple-system,BlinkMacSystemFont,”Segoe UI”,Helvetica,Arial,sans-serif}
.preview-wrap{max-width:900px;margin:0 auto}
.preview-header{padding:14px 18px;background:#ffffff;border:1px solid #e5e5e5;border-radius:8px;margin:0 0 18px 0;color:#666;font-size:13px}
Preview — this is how the embed will look inside your WordPress post.
#tdai-guide-mtp{background:#000000!important;color:#ffffff!important;font-family:-apple-system,BlinkMacSystemFont,”Segoe UI”,Helvetica,Arial,sans-serif!important;border:1px solid #1f1f1f!important;border-radius:14px!important;padding:0!important;margin:28px 0!important;overflow:hidden!important;position:relative!important;line-height:1.55!important;-webkit-font-smoothing:antialiased!important}
#tdai-guide-mtp *{box-sizing:border-box!important}
#tdai-guide-mtp hr,#tdai-guide-mtp p:empty,#tdai-guide-mtp del,#tdai-guide-mtp s{display:none!important}
#tdai-guide-mtp .tg-step-line{height:1px!important;background:#1f1f1f!important;border:0!important;margin:0!important;display:block!important}
#tdai-guide-mtp .tg-header{padding:28px 32px 22px 32px!important;border-bottom:1px solid #1f1f1f!important;background:#000000!important}
#tdai-guide-mtp .tg-eyebrow{font-size:11px!important;letter-spacing:2px!important;text-transform:uppercase!important;color:#888888!important;margin:0 0 10px 0!important;font-weight:600!important}
#tdai-guide-mtp .tg-title{font-size:26px!important;font-weight:700!important;color:#ffffff!important;margin:0!important;letter-spacing:-0.5px!important;line-height:1.25!important}
#tdai-guide-mtp .tg-subtitle{font-size:14px!important;color:#a0a0a0!important;margin:8px 0 0 0!important;font-weight:400!important}
#tdai-guide-mtp .tg-stage{position:relative!important;background:#000000!important;min-height:520px!important;padding:0!important}
#tdai-guide-mtp .tg-slides{position:relative!important;width:100%!important}
#tdai-guide-mtp .tg-slide{display:none!important;padding:32px 32px 24px 32px!important;animation:tg-fade 0.35s ease!important}
#tdai-guide-mtp .tg-slide.active{display:block!important}
@keyframes tg-fade{from{opacity:0;transform:translateY(6px)}to{opacity:1;transform:translateY(0)}}
#tdai-guide-mtp .tg-slide-num{font-size:11px!important;color:#666666!important;letter-spacing:2px!important;margin:0 0 14px 0!important;font-weight:600!important}
#tdai-guide-mtp .tg-slide-title{font-size:22px!important;font-weight:700!important;color:#ffffff!important;margin:0 0 8px 0!important;letter-spacing:-0.3px!important;line-height:1.3!important}
#tdai-guide-mtp .tg-slide-lead{font-size:14px!important;color:#a0a0a0!important;margin:0 0 22px 0!important;line-height:1.6!important}
#tdai-guide-mtp .tg-section{margin:0 0 18px 0!important}
#tdai-guide-mtp .tg-section h4{font-size:12px!important;letter-spacing:1.5px!important;text-transform:uppercase!important;color:#888888!important;margin:0 0 10px 0!important;font-weight:600!important}
#tdai-guide-mtp p{font-size:14px!important;color:#e0e0e0!important;margin:0 0 12px 0!important;line-height:1.6!important}
#tdai-guide-mtp ul{margin:0 0 14px 0!important;padding:0 0 0 18px!important;list-style:none!important}
#tdai-guide-mtp ul li{font-size:14px!important;color:#e0e0e0!important;margin:0 0 8px 0!important;line-height:1.55!important;position:relative!important;padding-left:14px!important}
#tdai-guide-mtp ul li::before{content:”›”!important;position:absolute!important;left:-4px!important;top:0!important;color:#ffffff!important;font-weight:700!important}
#tdai-guide-mtp code{background:#111111!important;color:#ffffff!important;padding:2px 6px!important;border-radius:4px!important;font-family:”SFMono-Regular”,Consolas,”Liberation Mono”,Menlo,monospace!important;font-size:12.5px!important;border:1px solid #1f1f1f!important}
#tdai-guide-mtp pre{background:#0a0a0a!important;color:#ffffff!important;padding:16px 18px!important;border-radius:8px!important;border:1px solid #1f1f1f!important;overflow-x:auto!important;margin:0 0 14px 0!important;font-family:”SFMono-Regular”,Consolas,”Liberation Mono”,Menlo,monospace!important;font-size:12.5px!important;line-height:1.6!important}
#tdai-guide-mtp pre code{background:transparent!important;border:0!important;padding:0!important;color:#ffffff!important;font-size:12.5px!important}
#tdai-guide-mtp .tg-comment{color:#666666!important}
#tdai-guide-mtp .tg-pyramid{display:flex!important;flex-direction:column!important;align-items:center!important;gap:6px!important;margin:8px 0 18px 0!important}
#tdai-guide-mtp .tg-layer{border:1px solid #2a2a2a!important;background:#0a0a0a!important;padding:12px 18px!important;border-radius:6px!important;color:#ffffff!important;font-size:13px!important;text-align:center!important;display:flex!important;flex-direction:column!important;gap:2px!important}
#tdai-guide-mtp .tg-layer-l3{width:55%!important;background:#1a1a1a!important;border-color:#3a3a3a!important}
#tdai-guide-mtp .tg-layer-l2{width:70%!important;background:#141414!important;border-color:#303030!important}
#tdai-guide-mtp .tg-layer-l1{width:85%!important;background:#0d0d0d!important;border-color:#262626!important}
#tdai-guide-mtp .tg-layer-l0{width:100%!important;background:#0a0a0a!important;border-color:#222222!important}
#tdai-guide-mtp .tg-layer-name{font-weight:700!important;font-size:13px!important;color:#ffffff!important;letter-spacing:0.3px!important}
#tdai-guide-mtp .tg-layer-desc{font-size:11.5px!important;color:#888888!important}
#tdai-guide-mtp .tg-table-wrap{overflow-x:auto!important;margin:0 0 14px 0!important;border:1px solid #1f1f1f!important;border-radius:8px!important}
#tdai-guide-mtp table{width:100%!important;border-collapse:collapse!important;font-size:12.5px!important;color:#ffffff!important;background:#000000!important}
#tdai-guide-mtp table th,#tdai-guide-mtp table td{padding:10px 12px!important;text-align:left!important;border-bottom:1px solid #1f1f1f!important;border-right:1px solid #1f1f1f!important;color:#e0e0e0!important}
#tdai-guide-mtp table th{background:#0a0a0a!important;color:#ffffff!important;font-weight:600!important;font-size:11.5px!important;letter-spacing:0.5px!important;text-transform:uppercase!important}
#tdai-guide-mtp table tr:last-child td{border-bottom:0!important}
#tdai-guide-mtp table th:last-child,#tdai-guide-mtp table td:last-child{border-right:0!important}
#tdai-guide-mtp .tg-grid-2{display:grid!important;grid-template-columns:1fr 1fr!important;gap:14px!important;margin:0 0 14px 0!important}
#tdai-guide-mtp .tg-card{border:1px solid #1f1f1f!important;background:#0a0a0a!important;border-radius:8px!important;padding:16px 18px!important}
#tdai-guide-mtp .tg-card h5{font-size:12px!important;letter-spacing:1px!important;text-transform:uppercase!important;color:#888888!important;margin:0 0 8px 0!important;font-weight:600!important}
#tdai-guide-mtp .tg-card p{font-size:13px!important;color:#e0e0e0!important;margin:0!important}
#tdai-guide-mtp .tg-controls{display:flex!important;align-items:center!important;justify-content:space-between!important;padding:18px 32px!important;border-top:1px solid #1f1f1f!important;background:#000000!important}
#tdai-guide-mtp .tg-dots{display:flex!important;gap:8px!important;align-items:center!important}
#tdai-guide-mtp .tg-dot{width:7px!important;height:7px!important;border-radius:50%!important;background:#2a2a2a!important;cursor:pointer!important;border:0!important;padding:0!important;transition:background 0.2s ease!important}
#tdai-guide-mtp .tg-dot.active{background:#ffffff!important}
#tdai-guide-mtp .tg-nav{display:flex!important;gap:10px!important;align-items:center!important}
#tdai-guide-mtp .tg-btn{background:#0a0a0a!important;color:#ffffff!important;border:1px solid #2a2a2a!important;width:38px!important;height:38px!important;border-radius:50%!important;cursor:pointer!important;display:flex!important;align-items:center!important;justify-content:center!important;font-size:18px!important;font-weight:400!important;line-height:1!important;padding:0!important;transition:all 0.2s ease!important;font-family:inherit!important}
#tdai-guide-mtp .tg-btn:hover{background:#ffffff!important;color:#000000!important;border-color:#ffffff!important}
#tdai-guide-mtp .tg-btn:disabled{opacity:0.3!important;cursor:not-allowed!important;background:#0a0a0a!important;color:#ffffff!important;border-color:#2a2a2a!important}
#tdai-guide-mtp .tg-counter{font-size:12px!important;color:#666666!important;letter-spacing:0.5px!important;min-width:40px!important;text-align:center!important}
#tdai-guide-mtp .tg-tagline{padding:18px 32px 22px 32px!important;border-top:1px solid #1f1f1f!important;background:#000000!important;text-align:center!important}
#tdai-guide-mtp .tg-tagline-text{font-size:11px!important;color:#666666!important;letter-spacing:1.5px!important;text-transform:uppercase!important;margin:0!important;font-weight:500!important}
#tdai-guide-mtp .tg-tagline-text strong{color:#ffffff!important;font-weight:700!important;letter-spacing:2px!important}
#tdai-guide-mtp .tg-callout{border-left:2px solid #ffffff!important;background:#0a0a0a!important;padding:12px 16px!important;border-radius:0 6px 6px 0!important;margin:0 0 14px 0!important;font-size:13px!important;color:#e0e0e0!important}
@media (max-width:640px){
#tdai-guide-mtp{margin:16px 0!important;border-radius:10px!important}
#tdai-guide-mtp .tg-header{padding:20px 18px 16px 18px!important}
#tdai-guide-mtp .tg-title{font-size:20px!important}
#tdai-guide-mtp .tg-subtitle{font-size:13px!important}
#tdai-guide-mtp .tg-stage{min-height:auto!important}
#tdai-guide-mtp .tg-slide{padding:22px 18px 18px 18px!important}
#tdai-guide-mtp .tg-slide-title{font-size:18px!important}
#tdai-guide-mtp .tg-slide-lead{font-size:13px!important}
#tdai-guide-mtp p,#tdai-guide-mtp ul li{font-size:13px!important}
#tdai-guide-mtp pre,#tdai-guide-mtp pre code{font-size:11.5px!important;padding:12px 14px!important}
#tdai-guide-mtp .tg-grid-2{grid-template-columns:1fr!important}
#tdai-guide-mtp .tg-controls{padding:14px 18px!important}
#tdai-guide-mtp .tg-btn{width:34px!important;height:34px!important;font-size:16px!important}
#tdai-guide-mtp .tg-tagline{padding:14px 18px 18px 18px!important}
#tdai-guide-mtp .tg-tagline-text{font-size:10px!important;letter-spacing:1.2px!important}
#tdai-guide-mtp .tg-layer-l3{width:75%!important}
#tdai-guide-mtp .tg-layer-l2{width:85%!important}
#tdai-guide-mtp .tg-layer-l1{width:95%!important}
#tdai-guide-mtp table th,#tdai-guide-mtp table td{padding:8px 10px!important;font-size:11.5px!important}
}
Open Source / Tencent
TencentDB Agent Memory
A quick-start guide to fully local, 4-tier long-term memory for AI agents.
01 / OVERVIEW
What is TencentDB Agent Memory?
An MIT-licensed memory system for AI agents that combines symbolic short-term memory with a 4-tier long-term memory pipeline. Runs fully local with zero external API dependencies.
Short-term memory
Offloads verbose tool logs to files and keeps a compact Mermaid task canvas in context.
Long-term memory
Distills conversations into a 4-tier semantic pyramid: L0 → L1 → L2 → L3.
Local backend
Defaults to SQLite + sqlite-vec. Tencent Cloud Vector Database (TCVDB) is optional.
Integrations
Ships as an OpenClaw plugin and a Hermes Agent Docker image.
02 / ARCHITECTURE
The 4-Tier Semantic Pyramid
Long-term memory is layered, not flat. Upper layers carry structure; lower layers preserve evidence.
L3 · PersonaUser profile (persona.md)
L2 · ScenarioScene blocks (Markdown)
L1 · AtomAtomic facts (JSONL)
L0 · ConversationRaw dialogue
Drill-down path: Persona → Scenario → Atom → Conversation. References use node_id and result_ref for deterministic traceback.
03 / SYMBOLIC SHORT-TERM
Mermaid task canvas + context offloading
Verbose intermediate logs are the largest token consumers in long tasks. The plugin offloads them to disk and keeps a high-density symbol graph in context.
How it works
Full tool logs are offloaded to refs/*.md under the data directory.
State transitions are encoded in Mermaid syntax inside a lightweight task canvas.
The agent reasons over the symbol graph, then greps a node_id to pull raw text.
Storage path on disk: ~/.openclaw/memory-tdai/. All artifacts are human-readable for white-box debugging.
04 / INSTALL
Install the OpenClaw plugin
Requires Node.js 22.16 or higher and an OpenClaw installation.
# Install the npm package as an OpenClaw plugin
openclaw plugins install @tencentdb-agent-memory/memory-tencentdb
openclaw gateway restart
Zero-config enable
Add the following to ~/.openclaw/openclaw.json to turn it on with default SQLite + sqlite-vec.
{
“memory-tencentdb”: {
“enabled”: true
}
}
05 / CONFIGURATION
Daily-tuning parameters
Every field has a sensible default. The most common knobs are listed below.
FieldDefaultDescription
storeBackendsqliteStorage backend
recall.strategyhybridkeyword / embedding / hybrid (RRF)
recall.maxResults5Items returned per recall
recall.timeoutMs5000Skip injection on timeout
pipeline.everyNConversations5L1 extraction every N turns
persona.triggerEveryN50Generate persona every N memories
offload.enabledfalseShort-term compression toggle
06 / SHORT-TERM COMPRESSION
Enable Mermaid offloading (v0.3.4+)
Three steps to turn on context offload for long-horizon tasks.
Step 1 · Enable offload in plugin config
{
“memory-tencentdb”: {
“config”: {
“offload”: { “enabled”: true }
}
}
}
Step 2 · Register the slot so OpenClaw routes offload requests
{
“plugins”: {
“slots”: {
“contextEngine”: “openclaw-context-offload”
}
}
}
Step 3 · Apply the runtime patch (once per OpenClaw install)
bash scripts/openclaw-after-tool-call-messages.patch.sh
07 / HERMES DOCKER
Run memory-enabled Hermes in one container
A single Docker image bundles Hermes Agent, the memory_tencentdb plugin, and the TDAI Memory Gateway.
# Build the image
docker build -f Dockerfile.hermes -t hermes-memory .
# Run the container (default model: DeepSeek-V3.2 on Tencent Cloud LKE)
docker run -d
–name hermes-memory
–restart unless-stopped
-p 8420:8420
-e MODEL_API_KEY=”your-api-key”
-e MODEL_BASE_URL=”https://api.lkeap.cloud.tencent.com/v1″
-e MODEL_NAME=”deepseek-v3.2″
-e MODEL_PROVIDER=”custom”
-v hermes_data:/opt/data
hermes-memory
# Health check
curl http://localhost:8420/health
Any OpenAI-compatible endpoint works through MODEL_PROVIDER=custom. Memory data persists in the hermes_data volume.
08 / AGENT TOOLS & RECALL
What the agent sees
Two tools are exposed to the agent during a session. Recall uses BM25 + vector + RRF fusion by default.
tdai_memory_search
Search across L1 Atoms, L2 Scenarios, and L3 Persona.
tdai_conversation_search
Search raw L0 Conversation history.
Retrieval defaults
Hybrid strategy: BM25 keyword + vector embedding, fused via Reciprocal Rank Fusion.
BM25 tokenizer supports Chinese (jieba) and English.
Returns 5 items per recall; 5000 ms timeout; on timeout it skips injection.
References include node_id and result_ref for traceback.
09 / BENCHMARKS
Reported gains with OpenClaw
Measured over continuous long-horizon sessions, not isolated turns. SWE-bench runs 50 consecutive tasks per session.
BenchmarkBaselineWith PluginΔ PassΔ Tokens
WideSearch33%50%+51.52%−61.38%
SWE-bench58.4%64.2%+9.93%−33.09%
AA-LCR44.0%47.5%+7.95%−30.98%
PersonaMem48%76%+59%—
Numbers come from Tencent’s own evaluations and reflect the integration with OpenClaw.
10 / RESOURCES
Where to go next
Documentation, source code, and community channels.
Source code
github.com/Tencent/TencentDB-Agent-Memory
npm package
@tencentdb-agent-memory/memory-tencentdb
License
MIT
Roadmap
Portable memory, automatic Skill generation, visual debugging dashboard.
‹
01 / 10
›
Curated by MARKTECHPOST · AI Research, Engineered for Builders
(function(){
var root = document.getElementById(‘tdai-guide-mtp’);
if(!root) return;
var slides = root.querySelectorAll(‘.tg-slide’);
var prev = root.querySelector(‘#tg-prev-mtp’);
var next = root.querySelector(‘#tg-next-mtp’);
var counter = root.querySelector(‘#tg-counter-mtp’);
var dotsWrap = root.querySelector(‘#tg-dots-mtp’);
var total = slides.length;
var current = 0;
function pad(n){return n