Context pollution and rot are probably more important than memory, because facts can usually be retrieved if the agent is good at following breadcrumbs.
What's also the biggest killer is code rot. Agents are particularly good at death by thousand cuts. They implement something poorly, or incorrectly, or introduce a bad pattern into the project. Then they continue to amplify that badness over time, as they continue to copy from it on subsequent work. It spreads like a virus.
Keeping these seeds out of the project is very difficult, and cleaning up the rot is very difficult. It also seems like a hard problem to solve because following the existing codebase is something that is good when the code is good, but bad when it is bad. So, seemingly, the solution means more thinking and evaluation for every change that is being made.
> They implement something poorly, or incorrectly, or introduce a bad pattern into the project. Then they continue to amplify that badness over time, as they continue to copy from it on subsequent work. It spreads like a virus.
> Keeping these seeds out of the project is very difficult, and cleaning up the rot is very difficult.
My "aha" moment was when I realized this goes for all spheres of life where this tech is/will be introduced.
It goes for all spheres of life, full stop. I’m not sure if agents struggle with this because they learned it from humans, or if they struggle with it because it’s a universally challenging problem, but it’s something we share with them.
ACM, that's the term that I'd been looking for - and your paper explains it clearly. At the end, most of LLM problems are context problems. Getting the correct knowledge into its context window without overpopulating it is the actual engineering effort for most agents. And the solution you present seems promising.
Both compaction with validation and predictive fetching are the way to go.
I do not want to write an implementation for this myself, and if Synap is that implementation, I'd like to ask you a few questions:
1. Does it work with context that's not just agent conversations, but rather documents?
2. Is it better than RAG on large dataset?
3. What does on-prem options look like?
1. Yes, works on docs, agent conversations, human-conversations from different sources (Slack, JIRA, etc.). We have connectors for some of these as well; so it is plug and play
2. conventional RAG recall accuracy is quite low (50-60%) and latency is pretty high (seconds). But worst is the precision; you end up context stuffing to get acceptable recall
3. We do offer on-prem deployments, but only on sizeable annual contracts
For me, if I put costs in my architecture, I'm limited heavily and that can easily change how memory is shaped dramatically.
The opposite, putting memory in architecture, is not true. Memory engineering first, then full scale in the system then costs considerations.
In addition, I believe this is future friendly. Because AI is advancing and getting smarter and cheaper everyday.
I couldn't find a guide on this so I share my basic thoughts.
Context drift on retries is easily the most annoying part of this setup. Locking down the tool payload schema first was the only thing that worked for us
Ive never read a paper cover to cover before but after wrestling with opus 5s english this paper is such a relief to read, its like my eyes has been washed off opus stink
Context pollution and rot are probably more important than memory, because facts can usually be retrieved if the agent is good at following breadcrumbs.
What's also the biggest killer is code rot. Agents are particularly good at death by thousand cuts. They implement something poorly, or incorrectly, or introduce a bad pattern into the project. Then they continue to amplify that badness over time, as they continue to copy from it on subsequent work. It spreads like a virus.
Keeping these seeds out of the project is very difficult, and cleaning up the rot is very difficult. It also seems like a hard problem to solve because following the existing codebase is something that is good when the code is good, but bad when it is bad. So, seemingly, the solution means more thinking and evaluation for every change that is being made.
Yes so regular human and agentic evaluation of the coding agent output, scoring it on specific criteria?
https://github.com/harness/harness-evals
> Then they continue to amplify that badness over time
Also, with "self-bias", models are also likely to grow new content into spots that match their subtle fingerprints from the past.
That might come at the expense of whatever corrected "we should avoid that and do this instead" alternative some human added for future architecture.
> They implement something poorly, or incorrectly, or introduce a bad pattern into the project. Then they continue to amplify that badness over time, as they continue to copy from it on subsequent work. It spreads like a virus.
> Keeping these seeds out of the project is very difficult, and cleaning up the rot is very difficult.
My "aha" moment was when I realized this goes for all spheres of life where this tech is/will be introduced.
It goes for all spheres of life, full stop. I’m not sure if agents struggle with this because they learned it from humans, or if they struggle with it because it’s a universally challenging problem, but it’s something we share with them.
The comment highlighted how LLMs exacerbate the issue by entrenching the preexisting issues.
Indeed, and reply to comment seemed a "yes and" -- As with humans.
It's curious how much of these could apply to either:
https://en.wikipedia.org/wiki/Reconstructive_memory
https://en.wikipedia.org/wiki/Misinformation_effect
And many mechanisms exacerbate issues by entrenching preexisting issues.
Truly. Doing this for coding agents is an interesting and different shaped problem.
Im wondering how silent information loss is detected later and what exactly the validation score measures.
ACM, that's the term that I'd been looking for - and your paper explains it clearly. At the end, most of LLM problems are context problems. Getting the correct knowledge into its context window without overpopulating it is the actual engineering effort for most agents. And the solution you present seems promising.
Both compaction with validation and predictive fetching are the way to go.
I do not want to write an implementation for this myself, and if Synap is that implementation, I'd like to ask you a few questions: 1. Does it work with context that's not just agent conversations, but rather documents? 2. Is it better than RAG on large dataset? 3. What does on-prem options look like?
Thanks Samyakk!
1. Yes, works on docs, agent conversations, human-conversations from different sources (Slack, JIRA, etc.). We have connectors for some of these as well; so it is plug and play 2. conventional RAG recall accuracy is quite low (50-60%) and latency is pretty high (seconds). But worst is the precision; you end up context stuffing to get acceptable recall 3. We do offer on-prem deployments, but only on sizeable annual contracts
I like to start with memory engineering then reach full system then reducing costs. This allows unlocking full potential of agents.
Interesting. Where can I read more about this?
I came up with this after building a few systems.
For me, if I put costs in my architecture, I'm limited heavily and that can easily change how memory is shaped dramatically. The opposite, putting memory in architecture, is not true. Memory engineering first, then full scale in the system then costs considerations.
In addition, I believe this is future friendly. Because AI is advancing and getting smarter and cheaper everyday.
I couldn't find a guide on this so I share my basic thoughts.
Context drift on retries is easily the most annoying part of this setup. Locking down the tool payload schema first was the only thing that worked for us
Nice, is there any harness that implements this approach?
I also wrote a shorter preview here: https://www.maximem.ai/blog/agentic-context-management-paper
your website isnt working
Apologies. Should be working now.
"I"
Ive never read a paper cover to cover before but after wrestling with opus 5s english this paper is such a relief to read, its like my eyes has been washed off opus stink
Haha! I am going to put this one up as a win! Thanks for reading! Hope you found it useful.
Yet, it is full of AI slop one-liners like: "The contest ahead is not over who stores the most data; it is over who manages context the best".