You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository was archived by the owner on Oct 4, 2026. It is now read-only.
Repository navigation
This repository was archived by the owner on Oct 4, 2026. It is now read-only.
[Feature Request] Swarm tools like "memory system" #17
Hi, as recently discussed on the OmO Slim repo I think this project could really benefit from a Swarm Tools like memory system which can store reasoning for features, things to keep in mind for the codebase etc.
I think the memory is the hive system itself and philosphy is that memory is only needed for each distinct feature that trickles down to each task. Avoiding creating monolithic memory which may cause bias or hallucinations? Do you have a link to the swarm tools suggestion for OmO/OmO-slim?
I think the memory is the hive system itself and philosphy is that memory is only needed for each distinct feature that trickles down to each task. Avoiding creating monolithic memory which may cause bias or hallucinations? Do you have a link to the swarm tools suggestion for OmO/OmO-slim?
hi @juztim regarding a comprehensive memory we still need investigate on how to make it work, most of memory solution for agent on the market right now are not working consistent enough on practical use, also if you want project based memory that will duplicate with the AGENTS.md, as AGENTS.md is project-based instruction, if you want all agent session in project follow certain instruction like what framework, lib, best practice for project it should be on AGENTS.md already, a seperate memory would cause duplicated. Maybe instead or seperate memory system, we should use AGENTS.md instead (All model are trained to follow AGENTS.md also, mostly AGENTS.md is not well maintained throughout the project run time, Hive actively update project AGENTS.md could be a better solution here)
Regarding Swarm, Hive and Swarm is similar in termology (Bee, Hive,..) but the appoarch is totally different as Swarm focus on paralell agent, autonomous. Hive focus on quality gates, documentation, human in the loop
GSD is workflow based human manually control workflow, GSD:
Human manually use command to change phase (New to learn the command and know where to use)
Use atomic commit
Milestone roadmap based
More Complex workflow
Spec kit-ish appoarch on init spec for a project and define big roadmap
Hive is Agent toolkit which mean agent control workflow human act as quality gates, Hive different:
Agent handle workflow by using tool (easier to use but can be human need to know how to interact with agent properly)
Use Git worktrees
No milestone, roadmap, just feature
Simple workflow only focus on working on a feature addition to existing project, feature is independent of each other, more like context driven development rather than spec-kit
Have you compared the two? I recently finished tweaking my nvim configuration, and now I'm thinking about enhancing my opencode. I took a look at the swarm documentation, which introduces many new concepts that seem promising (including things like the logic of ultimate_bug_scanner).
I agree with some of the ideas behind agent-hive, but I'm not very familiar with how opencode handles communication between agents itself. I'll take a closer look at the documentation and code here tomorrow.
I'm not sure if the statement above will cause the owner to misunderstand, since its goal isn't to make pair programming itself better, and my original intention was simply to find a posture with less friction in the collaboration between humans and agents. I believe these are not in conflict.
In my imagination, pre-embedding some reasonable breakpoints where humans might need to intervene is great, and being able to minimize the context overhead of communication transfers between agents would be even better. Let me give an example: in many of my architectural designs, an agent might produce a first version, then have a reviewer agent1 evaluate it, and then the two agents keep arguing back and forth until they produce a version, and when writing tests, a new agent needs to come in to design the test architecture. During this process, I also need to go back and reconsider whether the initial architectural design was reasonable enough.
Then I manually implemented communication between agents through copy-paste, guiding them to argue in a direction I'm interested in. This helps me determine whether there's any complexity I hadn't thought of that needs to be solved, so I can keep the code quality under control.
opencode's default subagent is actually already quite usable (I myself have a subagent for entropy increase assessment as a reviewer), but it doesn't support multi-turn conversations between agents, and this plugin below seems to touch on that.
At the same time, I also want to preserve the possibility of introducing various features in specific directions to strengthen the corresponding agents. For example, for entropy increase governance, I currently rely mainly on prompts, but in the future I could encapsulate more skills and static detection tools, including things like ultimate_bug_scanner, or for agents like explorer that need to perform local exploration, I'm not sure if you could let me try
opportunities for tools like this,
I know my thoughts are scattered right now, I'm just beginning to think through the workflow I envision and where the capability boundaries of such agent frameworks lie. If the owner is willing, I'd be happy to make whatever contributions I can to this project.
Hi @jensenojs , did you use Swarm, my experience of embed also memory (vector based) that they not working well on complex task, something like chatbot, workflow agents they work fine but when use with complex task like software development they not perform as well due to unexpected context retrieval, hard to organize context
The more organize and managed context window of the agent is, the smarter they are, that why the Hive stick with file-based context .hive context feed into agent, AGENTS.md and SKILL.md as psudo-memory, because that way we easily control what feed into agent context window
The direction that I want to take with Agent Hive is we're building prod-ready professional tool for software developer, the more knowledge the developer is, the more Hive help me leverage the productivity, the goal is not produce software fast but make the output more prod ready
That why the workflow is heavy human in the loop, context management based, audit-able everything, so simplify and go full agentic is not the goal here
Introduce embed memory based would shift the pattern instead we are doing human copilot along with agent it become more agentic workflow where agent doing everything human on the loop not in the loop, that not a bad thing too, Hive set out to leverage agentic coding via HITL and the human role in the Hive workflow would be a Software Architect/Designer with agent copilot.
If we going full agentic workflow with HOTL pattern I'm interested in another fork from agent hive instead maybe call it Agent-Nexus, already have a private repo for that currently prototyping based on https://code.claude.com/docs/en/agent-teams read through the link on how claude code do this you will know what I mean about goes full agentic and agent communication
Right now Hive did have some form of agent communication via task, spec, report with orchestrator as supervisor communicate with human, but no worker peer worker direct communication, as my vision for Hive it not need at the moment, instead trying to make Hive do everything, I would want do a fork from Hive where we goes full Agentic Orchestrate
Ah, I'm not sure how you understand "engineering ready," but my current experience is that even if a solution looks very simple, proving that it's problem-free is really quite troublesome. I later discovered that it only required discussing the worst-case scenarios across a dozen or so situations to see if they met expectations, but beforehand it was actually a mess. Regarding this problem, relying solely on (opencode's) subagents can't really solve it—they tend to want to rush into the next phase.
I agree with your work experience regarding embedded RAG actually not being very useful, but to avoid our discussion drifting off focus, please allow me to reframe the argument.
Let me first emphasize the key point of your argument, which I completely agree with: "the goal is not produce software fast but make the output more prod ready". The organizational approach based on file context is very restrained (I may mention this again later), but intuitively I don't think Agent-Nexus stands in opposition to "human in the loop". In the earliest stages, I personally engaged in red-blue exercises with an Agent, reviewing its proposals. Now I dispatch several red team Agents to challenge technical solutions, during which I prune away discussions that are actually meaningless, ensuring the focus of technical review stays on reasonable directions. Claude-code's Agent Teams might lack appropriate intervention mechanisms for humans during the thinking process? But if some conversation progress is documented through files, or if there are reasonable hooks in place, humans have suitable ways to intervene, stop, or even just pause discussions between Agents and keep up with their conversation.
Additionally, I haven't led a team before, but recently I've felt that orchestrating Agents is in some ways quite similar to being a Tech leader. When an Agent gets stuck or deviates from the technical focus, we need to correct course. When there's an important design decision or a difficult-to-trace bug, we may need to conduct multiple rounds of reviews.
I want to emphasize this based on your principle that "the goal is not produce software fast but make the output more prod ready"
We simply need to intervene elegantly in the process of Agents "getting things done" and consolidate consensus-based solutions. If I were to summarize the above viewpoint in one sentence, it would be that I believe an Agents Team can also (at least potentially) serve "human on the loop".
Furthermore, while the usefulness of code embedding remains debatable, it addresses a "small enough" problem: how to better search code (I've now replaced the built-in search tool with Tilth). Or, to use another example, perhaps tools like Beads can manage implementation tasks better than Markdown. It also addresses a "small enough" problem. I think the differences between frameworks should be grounded in philosophy, or perhaps in orchestration and organization. The remaining configurations might have a best practice as a default, just as your framework allows configuring skills for agents. However, the purpose of orchestration should still be to allow people to better control the pace.
Hi @jensenojs, thanks for your detailed reply it been a while.
My view on "engineering ready" would be on complex task and project, as Hive main purpose actually my personal tool, and most of my project is microservices, big scalable tenant system lot of abstraction and big codebase because of that autonomous agent actually didn't help much but it even slow down my code workflow, that why I making Hive here which specific solving that use case by balance workflow between human and agents, address right responsibility for each stakeholder (human for architecture, design, consultant) (agent for discovery, implementation, documentation).
I did lead 3-4 teams before and depend on the teams, the complexity of the project, we can choose the how much autonomous agent need to be to work best with the situation.
About beads or any memory system, I would like the "memory implementation" to be a alternative implementation so user can choose which approach they want beads or file-based markdown, because I don't think beads would work with my use case but it can work on your workflow and vice versa (actually file-based markdown despite simple but it will work for everyone and require no extra setup I would like markdown to stay as default "memory setting")
If you can help us with the design and implementation it would be great as I'm busy on next feature which is Hive Network (a good memory design may help with this as well)
I will think about the design more, for now Hive is stable enough and actually working good at the moment, instead trying to put everything into Hive to have "robust tool" we should spend more time on the design, I will also introduce a develop branch as well so we can experimental there
We have different emphases on "engineering ready", and I find this both interesting and important. This might be complementary in the future. When I have time, I'll also try to put together a small demo to see if my ideas above can actually translate into some kind of productivity. If we think of agent teams' attention and human attention as a kind of resource, then the scenario I imagined above might be quite useful in some cases that haven't been well solved yet (if there could be some mechanism to focus everyone on discussing and breaking down a complex problem together).
I absolutely agree with markdown as the default "memory setting". I've tried various tools just for fun or the comfort of the experience, and I know this somewhat contradicts Occam's Razor hhh. If I wanted these tools to become "official", it would require much more serious argumentation
Thank you for the explanation! I'll sync with you when I have some more interesting progress, and I'm also happy to try and contribute to the Hive series again. I think I can learn some potentially valuable things from here. @tctinh
Activity
[-][Feature Request] Beads like "memory system"[/-][+][Feature Request] Swarm tools like "memory system"[/+]imarshallwidjaja commentedon Jan 24, 2026
I think the memory is the hive system itself and philosphy is that memory is only needed for each distinct feature that trickles down to each task. Avoiding creating monolithic memory which may cause bias or hallucinations? Do you have a link to the swarm tools suggestion for OmO/OmO-slim?
juztim commentedon Jan 24, 2026
alvinunreal/oh-my-opencode-slim#4
this is the related issue
tctinh commentedon Jan 24, 2026
hi @juztim regarding a comprehensive memory we still need investigate on how to make it work, most of memory solution for agent on the market right now are not working consistent enough on practical use, also if you want project based memory that will duplicate with the AGENTS.md, as AGENTS.md is project-based instruction, if you want all agent session in project follow certain instruction like what framework, lib, best practice for project it should be on AGENTS.md already, a seperate memory would cause duplicated. Maybe instead or seperate memory system, we should use AGENTS.md instead (All model are trained to follow AGENTS.md also, mostly AGENTS.md is not well maintained throughout the project run time, Hive actively update project AGENTS.md could be a better solution here)
Regarding Swarm, Hive and Swarm is similar in termology (Bee, Hive,..) but the appoarch is totally different as Swarm focus on paralell agent, autonomous. Hive focus on quality gates, documentation, human in the loop
The project that close to Hive right now would be GSD https://github.com/glittercowboy/get-shit-done with the different in execution that are:
GSD is workflow based human manually control workflow, GSD:
Hive is Agent toolkit which mean agent control workflow human act as quality gates, Hive different:
jensenojs commentedon Feb 18, 2026
@juztim
Have you compared the two? I recently finished tweaking my nvim configuration, and now I'm thinking about enhancing my opencode. I took a look at the swarm documentation, which introduces many new concepts that seem promising (including things like the logic of ultimate_bug_scanner).
I agree with some of the ideas behind agent-hive, but I'm not very familiar with how opencode handles communication between agents itself. I'll take a closer look at the documentation and code here tomorrow.
I'm not sure if the statement above will cause the owner to misunderstand, since its goal isn't to make pair programming itself better, and my original intention was simply to find a posture with less friction in the collaboration between humans and agents. I believe these are not in conflict.
In my imagination, pre-embedding some reasonable breakpoints where humans might need to intervene is great, and being able to minimize the context overhead of communication transfers between agents would be even better. Let me give an example: in many of my architectural designs, an agent might produce a first version, then have a reviewer agent1 evaluate it, and then the two agents keep arguing back and forth until they produce a version, and when writing tests, a new agent needs to come in to design the test architecture. During this process, I also need to go back and reconsider whether the initial architectural design was reasonable enough.
Then I manually implemented communication between agents through copy-paste, guiding them to argue in a direction I'm interested in. This helps me determine whether there's any complexity I hadn't thought of that needs to be solved, so I can keep the code quality under control.
opencode's default subagent is actually already quite usable (I myself have a subagent for entropy increase assessment as a reviewer), but it doesn't support multi-turn conversations between agents, and this plugin below seems to touch on that.
At the same time, I also want to preserve the possibility of introducing various features in specific directions to strengthen the corresponding agents. For example, for entropy increase governance, I currently rely mainly on prompts, but in the future I could encapsulate more skills and static detection tools, including things like ultimate_bug_scanner, or for agents like explorer that need to perform local exploration, I'm not sure if you could let me try
opportunities for tools like this,
or for edit, we might want to try
I know my thoughts are scattered right now, I'm just beginning to think through the workflow I envision and where the capability boundaries of such agent frameworks lie. If the owner is willing, I'd be happy to make whatever contributions I can to this project.
tctinh commentedon Feb 23, 2026
Hi @jensenojs , did you use Swarm, my experience of embed also memory (vector based) that they not working well on complex task, something like chatbot, workflow agents they work fine but when use with complex task like software development they not perform as well due to unexpected context retrieval, hard to organize context
The more organize and managed context window of the agent is, the smarter they are, that why the Hive stick with file-based context .hive context feed into agent, AGENTS.md and SKILL.md as psudo-memory, because that way we easily control what feed into agent context window
The direction that I want to take with Agent Hive is we're building prod-ready professional tool for software developer, the more knowledge the developer is, the more Hive help me leverage the productivity, the goal is not produce software fast but make the output more prod ready
That why the workflow is heavy human in the loop, context management based, audit-able everything, so simplify and go full agentic is not the goal here
Introduce embed memory based would shift the pattern instead we are doing human copilot along with agent it become more agentic workflow where agent doing everything human on the loop not in the loop, that not a bad thing too, Hive set out to leverage agentic coding via HITL and the human role in the Hive workflow would be a Software Architect/Designer with agent copilot.
If we going full agentic workflow with HOTL pattern I'm interested in another fork from agent hive instead maybe call it Agent-Nexus, already have a private repo for that currently prototyping based on https://code.claude.com/docs/en/agent-teams read through the link on how claude code do this you will know what I mean about goes full agentic and agent communication
Right now Hive did have some form of agent communication via task, spec, report with orchestrator as supervisor communicate with human, but no worker peer worker direct communication, as my vision for Hive it not need at the moment, instead trying to make Hive do everything, I would want do a fork from Hive where we goes full Agentic Orchestrate
jensenojs commentedon Feb 28, 2026
Sorry for slow response! @tctinh
Ah, I'm not sure how you understand "engineering ready," but my current experience is that even if a solution looks very simple, proving that it's problem-free is really quite troublesome. I later discovered that it only required discussing the worst-case scenarios across a dozen or so situations to see if they met expectations, but beforehand it was actually a mess. Regarding this problem, relying solely on (opencode's) subagents can't really solve it—they tend to want to rush into the next phase.
I agree with your work experience regarding embedded RAG actually not being very useful, but to avoid our discussion drifting off focus, please allow me to reframe the argument.
Let me first emphasize the key point of your argument, which I completely agree with: "the goal is not produce software fast but make the output more prod ready". The organizational approach based on file context is very restrained (I may mention this again later), but intuitively I don't think Agent-Nexus stands in opposition to "human in the loop". In the earliest stages, I personally engaged in red-blue exercises with an Agent, reviewing its proposals. Now I dispatch several red team Agents to challenge technical solutions, during which I prune away discussions that are actually meaningless, ensuring the focus of technical review stays on reasonable directions. Claude-code's Agent Teams might lack appropriate intervention mechanisms for humans during the thinking process? But if some conversation progress is documented through files, or if there are reasonable hooks in place, humans have suitable ways to intervene, stop, or even just pause discussions between Agents and keep up with their conversation.
Additionally, I haven't led a team before, but recently I've felt that orchestrating Agents is in some ways quite similar to being a Tech leader. When an Agent gets stuck or deviates from the technical focus, we need to correct course. When there's an important design decision or a difficult-to-trace bug, we may need to conduct multiple rounds of reviews.
I want to emphasize this based on your principle that "the goal is not produce software fast but make the output more prod ready"
We simply need to intervene elegantly in the process of Agents "getting things done" and consolidate consensus-based solutions. If I were to summarize the above viewpoint in one sentence, it would be that I believe an Agents Team can also (at least potentially) serve "human on the loop".
jensenojs commentedon Feb 28, 2026
Furthermore, while the usefulness of code embedding remains debatable, it addresses a "small enough" problem: how to better search code (I've now replaced the built-in search tool with Tilth). Or, to use another example, perhaps tools like Beads can manage implementation tasks better than Markdown. It also addresses a "small enough" problem. I think the differences between frameworks should be grounded in philosophy, or perhaps in orchestration and organization. The remaining configurations might have a best practice as a default, just as your framework allows configuring skills for agents. However, the purpose of orchestration should still be to allow people to better control the pace.
tctinh commentedon Mar 2, 2026
Hi @jensenojs, thanks for your detailed reply it been a while.
My view on "engineering ready" would be on complex task and project, as Hive main purpose actually my personal tool, and most of my project is microservices, big scalable tenant system lot of abstraction and big codebase because of that autonomous agent actually didn't help much but it even slow down my code workflow, that why I making Hive here which specific solving that use case by balance workflow between human and agents, address right responsibility for each stakeholder (human for architecture, design, consultant) (agent for discovery, implementation, documentation).
I did lead 3-4 teams before and depend on the teams, the complexity of the project, we can choose the how much autonomous agent need to be to work best with the situation.
About beads or any memory system, I would like the "memory implementation" to be a alternative implementation so user can choose which approach they want beads or file-based markdown, because I don't think beads would work with my use case but it can work on your workflow and vice versa (actually file-based markdown despite simple but it will work for everyone and require no extra setup I would like markdown to stay as default "memory setting")
If you can help us with the design and implementation it would be great as I'm busy on next feature which is Hive Network (a good memory design may help with this as well)
I will think about the design more, for now Hive is stable enough and actually working good at the moment, instead trying to put everything into Hive to have "robust tool" we should spend more time on the design, I will also introduce a develop branch as well so we can experimental there
jensenojs commentedon Mar 2, 2026
We have different emphases on "engineering ready", and I find this both interesting and important. This might be complementary in the future. When I have time, I'll also try to put together a small demo to see if my ideas above can actually translate into some kind of productivity. If we think of agent teams' attention and human attention as a kind of resource, then the scenario I imagined above might be quite useful in some cases that haven't been well solved yet (if there could be some mechanism to focus everyone on discussing and breaking down a complex problem together).
I absolutely agree with markdown as the default "memory setting". I've tried various tools just for fun or the comfort of the experience, and I know this somewhat contradicts Occam's Razor hhh. If I wanted these tools to become "official", it would require much more serious argumentation
Thank you for the explanation! I'll sync with you when I have some more interesting progress, and I'm also happy to try and contribute to the Hive series again. I think I can learn some potentially valuable things from here. @tctinh