EverMind vs TencentDB Agent Memory: Architecture, Deployment, and Evaluation
EverMind vs TencentDB Agent Memory: Architecture, Deployment, and Evaluation
Compare EverOS and TencentDB Agent Memory on memory architecture, deployment, integrations, team controls, and published evaluation results.
EverMind研究人员
About 7 minutes to read

EverMind and TencentDB Agent Memory both help AI agents carry useful context beyond a single conversation. The practical difference is how that memory fits into your application: EverOS offers a managed memory service and a Markdown-first open-source runtime, while TencentDB documents a plugin-based memory stack and a Team Memory Beta built around shared assets.
The choice is not simply personalization versus collaboration, or accuracy versus token savings. Both projects address several parts of the memory lifecycle. Compare the deployment you will operate, the information your agents need to retain, and the way you will test recall in your own workflow.
Key takeaways
• EverOS supports managed Cloud and self-hosted deployment, with user memory and agent experience represented separately.
• TencentDB's plugin path and Team Memory Beta are different deployment models; their architecture should not be collapsed into one local SQLite description.
• Both projects describe layered information and selective retrieval. The label “memory OS” does not establish a performance winner.
• TencentDB publishes accuracy and task-performance results as well as token savings; those results use different tests from the EverOS figures discussed here.
• Start with ownership, integration, and operating requirements, then validate the chosen system on representative tasks.
What are the two systems designed to do?
At EverMind, we build EverOS as a persistent memory foundation for agentic applications. It retains user context and execution experience so a later interaction can retrieve useful information instead of depending on the previous conversation remaining open.
User profiles and episodes capture personal context and meaningful interactions. Agent Cases and Skills capture task experience and reusable approaches. Applications choose the owner and scope that belong to the current workflow.
TencentDB Agent Memory's official project documentation describes both conversation memory and reusable team assets. Its Team Memory Beta combines Chat Memory, Skills, Wiki, and CodeGraph with a Hub for reviewing, sharing, and assigning those assets to agents.
These goals overlap. A personal assistant needs correct user context; a team of coding agents needs shared decisions and reusable procedures. Either use case benefits from explicit ownership, relevant retrieval, and a way to correct outdated information.
How do their architectures compare?
EverOS OSS uses Markdown as the source of truth, with SQLite for state and LanceDB for retrieval indexes. Its open-source architecture makes the memory files readable independently of those derived components. EverOS Cloud provides managed memory infrastructure for applications that do not want to operate the service themselves.
TencentDB's plugin documentation describes symbolic short-term memory alongside a long-term hierarchy: L0 Conversation, L1 Atom, L2 Scenario, and L3 Persona. The local plugin path defaults to SQLite plus sqlite-vec. The Team Memory Beta adds a separate Memory Core, Memory Hub, and Proxy deployment around reusable assets and team controls.
Dimension | EverOS | TencentDB Agent Memory |
|---|---|---|
Memory organization | User profiles and episodes; agent Cases and Skills | L0–L3 conversation hierarchy; Team Beta adds shared assets |
Storage | OSS: Markdown + SQLite + LanceDB | Plugin: SQLite + sqlite-vec; Team Beta uses a separate service stack |
Retrieval | Keyword, vector, hybrid, and additional documented retrieval paths | BM25 + vector + RRF; scoped access to team assets |
Experience reuse | Agent Cases distilled into reusable Skills | Team Beta manages Skills, Wiki, and CodeGraph |
Operating model | Managed Cloud or self-hosted OSS | Self-operated plugin or Team Memory Beta services |
The important distinction is not that one project has a lifecycle and the other only has a pipeline. Both describe capture, organization, retrieval, and reuse. Their components, operating models, and integration surfaces are what determine how they fit a particular application.
Retrieval should return useful evidence, not every saved detail
EverOS exposes keyword, vector, and hybrid search, with additional documented retrieval paths for more complex queries. The application supplies the intended identity and scope, retrieves relevant memory, and uses that context in its model request.
TencentDB describes layered retrieval with BM25, vector search, and reciprocal rank fusion. Its Team Memory documentation also describes permission-scoped asset access and limits on returned context. L2 and L3 can provide higher-level context, with lower layers available for more specific evidence.
In either system, retrieval quality depends on what was stored and what the query asks for. An external memory foundation can support portable context across agents, but each client still needs the correct integration and access boundaries.
What do the published benchmarks show?
EverMind's published memory benchmark results include 93.05% overall accuracy on LoCoMo and 83.00% on LongMemEval. TencentDB's plugin documentation reports PersonaMem accuracy improving from 48% to 76%, plus task-success and token-usage results from OpenClaw tests.
The published plugin benchmarks attribute the 61.38% token reduction to a WideSearch comparison. That is a result from a particular workload and setup, not a universal reduction for every request or a substitute for an accuracy measurement.
Reported evaluation | Published result | What it measures |
|---|---|---|
EverOS: LoCoMo | 93.05% overall accuracy | Recall in the cited long-conversation evaluation |
EverOS: LongMemEval | 83.00% accuracy | Performance in the cited long-term memory evaluation |
TencentDB: PersonaMem | 48% baseline → 76% with memory | User-information recall in its reported plugin setup |
TencentDB: WideSearch | 61.38% fewer tokens | Token usage in the reported OpenClaw task comparison |
These figures measure different outcomes under different conditions. A LoCoMo percentage and a PersonaMem percentage do not establish a head-to-head ranking. Token savings also need to be read alongside task success, latency, and the cost of processing and retrieving memory.
For a fair comparison, use the same data, questions, answer model, and scoring rules. Our memory evaluation framework explains why consistent methodology matters. In an application trial, include corrections, cross-session recall, temporal changes, and unrelated-user checks rather than testing only whether an agent can repeat a recently saved fact.
How do deployment and integrations differ?
EverOS gives developers a choice between managed Cloud and a self-hosted runtime. The self-hosted stack uses readable memory files and embedded storage components. Cloud lets the team consume the memory service without running that infrastructure.
EverOS also provides API and SDK access for custom applications. Its MCP integration documents connection paths for clients including Claude, ChatGPT, Claude Code, Cursor, and Codex. Configure each client for the intended memory space, owner, application, and project.
TencentDB's plugin path integrates memory into supported agent environments. Its Team Memory Beta instead documents a service stack and Proxy-based connection model. Choosing between those paths changes what you deploy, how you configure model access, and how you manage team assets.
Self-hosting is an operating decision, not proof that every processing step stays on one machine. Review the extraction and embedding providers used by the deployment. The application should route information according to its actual configuration, whether the memory store is local or managed.
How should teams compare sharing and control?
Sharing works best when it preserves context without making every user's information available to every agent. Decide who owns a memory, which project it belongs to, and which workflows can retrieve it.
EverOS separates user-owned profiles and episodes from agent-owned Cases and Skills. Its shared-memory controls use owner and application scope as part of retrieval. Your application pairs those boundaries with its authorization policy.
TencentDB's Team Memory Beta describes private, team, and restricted visibility, together with ownership, review, and agent asset bindings. Those are useful controls for a shared asset workflow; they are different from simply enabling the local conversation-memory plugin.
Compare the controls your team will actually use. A product assistant may need precise per-user isolation and a correction interface. An engineering team may also need a review process for shared Skills or code knowledge. The right deployment makes those responsibilities clear to its operators.
Which system fits your workflow?
Choose EverOS when you want a managed memory backend or a Markdown-first runtime you can operate yourself, alongside documented APIs and client integrations. It is a practical fit for developers building persistent assistants, customer-facing agents, and workflows that reuse context across sessions and tools.
TencentDB's documented plugin and Team Memory paths address different operating needs. Evaluate the plugin as an integration into the supported agent environment, and evaluate Team Memory Beta as a service-and-asset workflow. Do not assign the Beta's team controls to an older plugin setup by assumption.
Requirement | EverOS fit | TencentDB fit |
|---|---|---|
Managed infrastructure | EverOS Cloud operates the memory service | The cited setup guides describe services you operate |
User and agent experience | Profiles, episodes, Cases, and Skills | Layered Chat Memory; Skills in Team Memory Beta |
Team asset workflow | Integrate memory scopes into your application's controls | Team Beta adds Hub review, sharing, and agent bindings |
Self-operated runtime | Markdown-first OSS with embedded indexes | Plugin storage or the separate Team Memory service stack |
Before a wider rollout, run one complete memory loop: save useful context, allow processing to finish, retrieve it in a new session, correct it, and test again. The agent memory integration guide provides a broader view of that workflow. A successful trial shows that the right information reaches the right agent after it changes.
Frequently asked questions
Is EverOS only for personalization while TencentDB is only for teams?
No. EverOS includes user memory and retained agent experience, while TencentDB documents conversation memory and team assets. Compare their actual deployment and control paths rather than assigning each project a single use case.
Is TencentDB Agent Memory only a local SQLite pipeline?
That describes part of its plugin path, not the whole current project. The official repository also documents Team Memory Beta with Memory Core, Memory Hub, and Proxy services.
Do published scores prove which system will work better in production?
Published results describe performance in their stated evaluation setups. Use a matched test of your own data and workflow to compare recall, task success, latency, and total cost.
Build around the memory your agents need
Start with the information worth retaining and the operating model your team can support. Then test how the system captures, retrieves, corrects, and shares that information across real tasks. Explore EverOS when you want that memory foundation as a managed service or a self-hosted, readable runtime.
EverMind and TencentDB Agent Memory both help AI agents carry useful context beyond a single conversation. The practical difference is how that memory fits into your application: EverOS offers a managed memory service and a Markdown-first open-source runtime, while TencentDB documents a plugin-based memory stack and a Team Memory Beta built around shared assets.
The choice is not simply personalization versus collaboration, or accuracy versus token savings. Both projects address several parts of the memory lifecycle. Compare the deployment you will operate, the information your agents need to retain, and the way you will test recall in your own workflow.
Key takeaways
• EverOS supports managed Cloud and self-hosted deployment, with user memory and agent experience represented separately.
• TencentDB's plugin path and Team Memory Beta are different deployment models; their architecture should not be collapsed into one local SQLite description.
• Both projects describe layered information and selective retrieval. The label “memory OS” does not establish a performance winner.
• TencentDB publishes accuracy and task-performance results as well as token savings; those results use different tests from the EverOS figures discussed here.
• Start with ownership, integration, and operating requirements, then validate the chosen system on representative tasks.
What are the two systems designed to do?
At EverMind, we build EverOS as a persistent memory foundation for agentic applications. It retains user context and execution experience so a later interaction can retrieve useful information instead of depending on the previous conversation remaining open.
User profiles and episodes capture personal context and meaningful interactions. Agent Cases and Skills capture task experience and reusable approaches. Applications choose the owner and scope that belong to the current workflow.
TencentDB Agent Memory's official project documentation describes both conversation memory and reusable team assets. Its Team Memory Beta combines Chat Memory, Skills, Wiki, and CodeGraph with a Hub for reviewing, sharing, and assigning those assets to agents.
These goals overlap. A personal assistant needs correct user context; a team of coding agents needs shared decisions and reusable procedures. Either use case benefits from explicit ownership, relevant retrieval, and a way to correct outdated information.
How do their architectures compare?
EverOS OSS uses Markdown as the source of truth, with SQLite for state and LanceDB for retrieval indexes. Its open-source architecture makes the memory files readable independently of those derived components. EverOS Cloud provides managed memory infrastructure for applications that do not want to operate the service themselves.
TencentDB's plugin documentation describes symbolic short-term memory alongside a long-term hierarchy: L0 Conversation, L1 Atom, L2 Scenario, and L3 Persona. The local plugin path defaults to SQLite plus sqlite-vec. The Team Memory Beta adds a separate Memory Core, Memory Hub, and Proxy deployment around reusable assets and team controls.
Dimension | EverOS | TencentDB Agent Memory |
|---|---|---|
Memory organization | User profiles and episodes; agent Cases and Skills | L0–L3 conversation hierarchy; Team Beta adds shared assets |
Storage | OSS: Markdown + SQLite + LanceDB | Plugin: SQLite + sqlite-vec; Team Beta uses a separate service stack |
Retrieval | Keyword, vector, hybrid, and additional documented retrieval paths | BM25 + vector + RRF; scoped access to team assets |
Experience reuse | Agent Cases distilled into reusable Skills | Team Beta manages Skills, Wiki, and CodeGraph |
Operating model | Managed Cloud or self-hosted OSS | Self-operated plugin or Team Memory Beta services |
The important distinction is not that one project has a lifecycle and the other only has a pipeline. Both describe capture, organization, retrieval, and reuse. Their components, operating models, and integration surfaces are what determine how they fit a particular application.
Retrieval should return useful evidence, not every saved detail
EverOS exposes keyword, vector, and hybrid search, with additional documented retrieval paths for more complex queries. The application supplies the intended identity and scope, retrieves relevant memory, and uses that context in its model request.
TencentDB describes layered retrieval with BM25, vector search, and reciprocal rank fusion. Its Team Memory documentation also describes permission-scoped asset access and limits on returned context. L2 and L3 can provide higher-level context, with lower layers available for more specific evidence.
In either system, retrieval quality depends on what was stored and what the query asks for. An external memory foundation can support portable context across agents, but each client still needs the correct integration and access boundaries.
What do the published benchmarks show?
EverMind's published memory benchmark results include 93.05% overall accuracy on LoCoMo and 83.00% on LongMemEval. TencentDB's plugin documentation reports PersonaMem accuracy improving from 48% to 76%, plus task-success and token-usage results from OpenClaw tests.
The published plugin benchmarks attribute the 61.38% token reduction to a WideSearch comparison. That is a result from a particular workload and setup, not a universal reduction for every request or a substitute for an accuracy measurement.
Reported evaluation | Published result | What it measures |
|---|---|---|
EverOS: LoCoMo | 93.05% overall accuracy | Recall in the cited long-conversation evaluation |
EverOS: LongMemEval | 83.00% accuracy | Performance in the cited long-term memory evaluation |
TencentDB: PersonaMem | 48% baseline → 76% with memory | User-information recall in its reported plugin setup |
TencentDB: WideSearch | 61.38% fewer tokens | Token usage in the reported OpenClaw task comparison |
These figures measure different outcomes under different conditions. A LoCoMo percentage and a PersonaMem percentage do not establish a head-to-head ranking. Token savings also need to be read alongside task success, latency, and the cost of processing and retrieving memory.
For a fair comparison, use the same data, questions, answer model, and scoring rules. Our memory evaluation framework explains why consistent methodology matters. In an application trial, include corrections, cross-session recall, temporal changes, and unrelated-user checks rather than testing only whether an agent can repeat a recently saved fact.
How do deployment and integrations differ?
EverOS gives developers a choice between managed Cloud and a self-hosted runtime. The self-hosted stack uses readable memory files and embedded storage components. Cloud lets the team consume the memory service without running that infrastructure.
EverOS also provides API and SDK access for custom applications. Its MCP integration documents connection paths for clients including Claude, ChatGPT, Claude Code, Cursor, and Codex. Configure each client for the intended memory space, owner, application, and project.
TencentDB's plugin path integrates memory into supported agent environments. Its Team Memory Beta instead documents a service stack and Proxy-based connection model. Choosing between those paths changes what you deploy, how you configure model access, and how you manage team assets.
Self-hosting is an operating decision, not proof that every processing step stays on one machine. Review the extraction and embedding providers used by the deployment. The application should route information according to its actual configuration, whether the memory store is local or managed.
How should teams compare sharing and control?
Sharing works best when it preserves context without making every user's information available to every agent. Decide who owns a memory, which project it belongs to, and which workflows can retrieve it.
EverOS separates user-owned profiles and episodes from agent-owned Cases and Skills. Its shared-memory controls use owner and application scope as part of retrieval. Your application pairs those boundaries with its authorization policy.
TencentDB's Team Memory Beta describes private, team, and restricted visibility, together with ownership, review, and agent asset bindings. Those are useful controls for a shared asset workflow; they are different from simply enabling the local conversation-memory plugin.
Compare the controls your team will actually use. A product assistant may need precise per-user isolation and a correction interface. An engineering team may also need a review process for shared Skills or code knowledge. The right deployment makes those responsibilities clear to its operators.
Which system fits your workflow?
Choose EverOS when you want a managed memory backend or a Markdown-first runtime you can operate yourself, alongside documented APIs and client integrations. It is a practical fit for developers building persistent assistants, customer-facing agents, and workflows that reuse context across sessions and tools.
TencentDB's documented plugin and Team Memory paths address different operating needs. Evaluate the plugin as an integration into the supported agent environment, and evaluate Team Memory Beta as a service-and-asset workflow. Do not assign the Beta's team controls to an older plugin setup by assumption.
Requirement | EverOS fit | TencentDB fit |
|---|---|---|
Managed infrastructure | EverOS Cloud operates the memory service | The cited setup guides describe services you operate |
User and agent experience | Profiles, episodes, Cases, and Skills | Layered Chat Memory; Skills in Team Memory Beta |
Team asset workflow | Integrate memory scopes into your application's controls | Team Beta adds Hub review, sharing, and agent bindings |
Self-operated runtime | Markdown-first OSS with embedded indexes | Plugin storage or the separate Team Memory service stack |
Before a wider rollout, run one complete memory loop: save useful context, allow processing to finish, retrieve it in a new session, correct it, and test again. The agent memory integration guide provides a broader view of that workflow. A successful trial shows that the right information reaches the right agent after it changes.
Frequently asked questions
Is EverOS only for personalization while TencentDB is only for teams?
No. EverOS includes user memory and retained agent experience, while TencentDB documents conversation memory and team assets. Compare their actual deployment and control paths rather than assigning each project a single use case.
Is TencentDB Agent Memory only a local SQLite pipeline?
That describes part of its plugin path, not the whole current project. The official repository also documents Team Memory Beta with Memory Core, Memory Hub, and Proxy services.
Do published scores prove which system will work better in production?
Published results describe performance in their stated evaluation setups. Use a matched test of your own data and workflow to compare recall, task success, latency, and total cost.
Build around the memory your agents need
Start with the information worth retaining and the operating model your team can support. Then test how the system captures, retrieves, corrects, and shares that information across real tasks. Explore EverOS when you want that memory foundation as a managed service or a self-hosted, readable runtime.
您可能还喜欢这些
相关

Multi-round retrieval: letting the model decide when to stop searching
multi-round retrieval,agentic retrieval,fixed top-k,injection budget,core selection,EverOS,agent memory,LoCoMo,LongMemEval,EverMemBench,SubtleMemory,multi-hop retrieval,retrieval accuracy,prompt token cost

Do public SKILL.md files actually make agents better?
SkillCorpus,SKILL.md,agent skills,skill curation,skill retrieval,LLM agents,SkillsBench,GDPVal,agent harness

CRAFT: learning how to fuse video tokens, not just which to drop
CRAFT,video token compression,vision-language models,video VLM,KV cache,prefill cost,token merging,token pruning,temporal reasoning

Self-evolving agents have a measurement problem
self-evolving agents,agent harness,HarnessBank,credit assignment,LLM agents,agent evaluation,harness optimization,significance testing
EverMind vs TencentDB Agent Memory: Architecture, Deployment, and Evaluation
Compare EverOS and TencentDB Agent Memory on memory architecture, deployment, integrations, team controls, and published evaluation results.
EverMind研究人员
About 7 minutes to read


