JoyJoy

[JoyJoy @YuyunDeng]

Most AI agents still have a memory problem.

They can chat for hours, solve hard tasks, and appear “context-aware” — then you open a new window, switch chats, change platforms, or come back tomorrow… and suddenly the agent has forgotten the important parts.

Scope Recall is our answer to that.

Scope Recall is a standalone memory provider for Hermes Agent, built for agents that need memory to be durable, inspectable, scoped, and safe.

It gives Hermes a real local-first memory layer:

• current-turn semantic recall
• durable cross-chat / cross-window facts
• local scratch isolation, so temporary chat noise does not leak everywhere
• SQLite as the auditable source of truth
• LanceDB or SQLite vector companions for semantic retrieval
• hybrid lexical + BM25 + vector retrieval with RRF ranking
• entity probing and related-memory exploration
• compact task-relevant context generation
• profile/context surfaces for long-running agents
• trust feedback, inspect, explain, benchmark, stats, export, merge, update, and forget tools
• secret-safe credential indexes that point to your vault without storing plaintext secrets
• optional cross-platform identity mapping for Telegram, CLI, Feishu, and other gateway contexts
• governance and hygiene workflows for long-lived memory stores
• Experience Kernel playbooks for reusable procedural memory and preflight guidance

The design principle is simple:

SQLite is the truth.
Vectors are rebuildable companions.
Durable facts can travel deliberately.
Temporary scratch stays local.
Secrets stay out of memory.
Recall is based on the current query — not stale context from the previous topic.

This matters because “memory” is not just about remembering more.

Bad memory makes agents worse:
they leak context between tasks, preserve wrong assumptions, recall irrelevant history, store private data too casually, and become harder to audit over time.

Scope Recall is built around responsible memory:
local-first, inspectable, scoped, recoverable, and explicit about what should be durable vs. what should remain temporary.

For solo agents, it makes Hermes feel like it finally remembers across windows.

For multi-agent systems, it gives every agent a safer boundary between shared durable knowledge and local working context.

And Scope Recall is not limited to Hermes: an OpenClaw sibling implementation brings the same local-first, scope-aware, SQLite-backed memory architecture to OpenClaw agents as well.

For operators, Scope Recall provides real visibility:
inspect what was stored, explain why something was recalled, benchmark retrieval behavior, export scoped rows, repair vector companions, and clean up memory drift without treating the vector index as magic.

We built Scope Recall because agent memory should not be a black box.

It should be something you can query, audit, repair, and trust.

Scope Recall turns agent memory from “notes in a prompt” into an actual memory system.

Local-first.
SQLite-backed.
Vector-assisted.
Scope-aware.
Governance-ready.

For agents that actually remember.

Hermes version:
https://t.co/HNayVS6yKn

OpenClaw version:
https://t.co/IfYchIIrA6

PyPI:
https://t.co/glHlYokX1s

[JoyJoy @YuyunDeng]

卧槽啊！今天试了下和GPT语音对话，完全感觉对面是个中文不太标准的真人，说话和真人简直一模一样，一点点都不AI了，chatgpt语音感觉比豆包强多了，文字沟通时候的那些口癖例如稳稳接一刀完全没了，还有思维方式完全感觉不出是AI在输出文字和声音了

[JoyJoy @YuyunDeng]

@yangye0421 昨天应该恢复了吧？前面2个月我都被路由到5.5 mini,昨天恢复到pro了

[JoyJoy @YuyunDeng]

@alswl 国内不管是kimi还是glm的，我就没见过量大的，哎，他们能有codex那个额度，200美金一月会员我也愿意定一个

[JoyJoy @YuyunDeng]

@1345ccb @Fomsky_Wei 昨天都还路由到5.5mini,今天竟然不路由了

[JoyJoy @YuyunDeng]

@thsottiaux Thanks, Tibo, for the reset!!!!!! I noticed it reset again when I still had 30% left. Thank you so much, Tibo!

[JoyJoy @YuyunDeng]

@jun_song 因为中国这边服务器涨价太厉害了，几倍的利润啊

[JoyJoy @YuyunDeng]

Including Pro users, please add the 5-hour limit across the board, but please don’t quietly reduce the total weekly quota.

The weekly allowance already gets used up way too quickly now. I remember back in June and July, when I was using GPT-5.5, I could hardly ever hit the 5-hour limit, and my weekly quota would usually last all the way until the final day.

But now... if I use it the same way I did before, I can burn through my entire weekly quota in two days at most.

[JoyJoy @YuyunDeng]

@Da7_Tech i think 5.6 sol>grok4.6>=deepseek😂

[JoyJoy @YuyunDeng]

@Shenxiao123971 我是真心想遇到个这种实心眼的姑娘，就是遇不到啊，哎，真是缘分没到，你是遇到了自己放弃了

[JoyJoy @YuyunDeng]

I haven’t used your product, so I won’t comment on its functionality. But compared with the official Hermes website, your frontend design is honestly terrible—I’m a designer myself. Speaking purely in terms of aesthetics and design, your product looks like a toad in the dirt next to Hermès’s website, which is like a swan in the sky.

I get the feeling that you’re trying to piggyback on Hermes’s buzz and traffic, which makes you look like a complete clown.

Hermes does fall short in some areas, but for now, among similar AI agents, its overall user experience is genuinely quite good.

[JoyJoy @YuyunDeng]

@AI_Community_9 到底怎么做到的啊，我也是5090，怎么我用sglang或者lm studio都最高只能到50-70

[JoyJoy @YuyunDeng]

@thsottiaux https://t.co/wI1tHMMdJU

[叶客 @Yyylllyly]

我不是Token神

[JoyJoy @YuyunDeng]

内部人员刚刚透露了Ox Alpha 到底是什么模型了，哈哈，真的想不到，想不到啊

[JoyJoy @YuyunDeng]

@Nominatiivi @songkeys @grok 此方法是否有效，是否有更好的建议

[JoyJoy @YuyunDeng]

@ATongtian69578 @CuiMao 我们可以做纹身~贴纸那种，左千问，右kimi

[JoyJoy @YuyunDeng]

@brucexu_eth 我也是，从OPENCLAW时代到hermes用到现在，这个在我用来比codex现在更顺手

[JoyJoy @YuyunDeng]

@florin_dev r u sure?

[JoyJoy @YuyunDeng]

@yiyimuning @CuiMao @DarioAmodei 哈哈哈哈哈哈哈哈

[JoyJoy @YuyunDeng]

@tualatrix 我是heavy套餐。。最多两天就用完了。。我用5.6 sol指挥3个grok4.6同时去做事情，额度消耗的真的真的非常快，现在我把送的cursor也拉了cli出来做一个帮手，一起用

[JoyJoy @YuyunDeng]

I disagree with this view.

A2A and bot-to-bot communication are extremely useful to me right now, especially A2A.

Before this, I had been experimenting on my own with ways to let agents communicate with each other. But because of the limitations of my coding skills, the results were never quite as good as I wanted.

At one point, I even created Telegram group chats for my agents so they could communicate with each other and assign tasks. It worked to some extent, but the overall effectiveness was still fairly limited.

Now, with the A2A and bot features, everything has changed.

Each of my instances can now function as a small team with its own specialized role, and I can use A2A to distribute tasks between those teams. It works extremely well.

Previously, I used GPT-5.6 Sol directly for most of my work, and I was constantly running into usage limits.

Now, through A2A, I use one GPT-5.6 Sol instance as the lead commander, coordinating several Grok 4.6 instances together with a DeepSeek V4 Flash bot to complete the same project collaboratively.

The time required has dropped dramatically, and more importantly, this setup allows GPT-5.6 Sol to make much better use of what it is best at: being the brain and commander of the entire system, rather than spending its valuable tokens doing every task itself.

And the most important part?

After running this setup for 24 hours, my Codex weekly quota had decreased by only 3%.

The shared workspace issue is also very easy to solve.

I simply deployed PostgreSQL + pgvector as a shared knowledge base and shared memory layer for all of my instances.

Then, each individual instance uses a lightweight memory plugin that I developed myself for its own private memory.

The whole setup works incredibly well.

Overall, Hermes has now reached an almost perfect state for my workflow.

It has helped me solve a huge number of real-world work problems, and in terms of both smoothness and stability, it is far beyond what I previously experienced with OpenClaw.