▲ 208 points
back
90 comments
Instead of compaction, has anyone seen a successful implementation of pruning? That is, the agent looks at the conversation history and removes any low-value messages.
For example, sometimes context will be taken up by a side tangent, tool call outputs, or low-value codebase exploration.
Much of the time, I prefer to preserve the history of my conversation instead of summarizing it. I find summarized conversations lead to more frustrating future chats because the LLM misses intent and or context. (Or, the presence of paragraphs and paragraphs of LLM output makes the next token predictor dumber? Unsure.)
It is kinda a combination of the two:
https://github.com/spott/pi-task-compaction
I’m still playing with it, but it essentially has the model define a region, and mark it with begin_task, and end_task. End_task also requires a fairly hefty summery of the entire region.
It then pulls that entire region and replaces it with the summary. If the model wants it can actually look into the output that was pruned from the conversation, but in practice the summaries have been good enough it hasn’t needed to.
I’ve been using it with 5.6 sol, and even really long sessions have like 6% context used at the end.
I think there are a lot of strategies that will open up when costs come down but right now you take a hit on cache rate and thus costs every time you do anything other than wait until the last minute to compact
I built a harness for my own use that allows the agent to fork its own history. So for example it can “compact from” a specific item, replacing a branch with a summary; or do an “excursion”, basically a temporary branch, which is like a subagent but inherits context.
Sounds cool and it does make sensible decisions optically but I haven’t been able to prove that it is meaningfully better than normal compaction. Building harnesses that do interesting things is a lot easier than building more effective harnesses, I guess.
I'm very interested in this too. I feel like when my Claude session compacts I immediately have to re-alert it to critical aspects of the task, but it should be possible even to have a secondary low-skill agent crawl over the whole context window semi-continuously and nominate pieces for removal or summarization. Or at a higher level, have something able to decompose, like hey I realised we're actually working on three distinct aspects of this problem, let's split the context into those three and I'll just retain a high level summary of what's in the other ones, enough to know if I should pass off control to one of the others for a particular ask.
I feel like this is kind of subagents, but it doesn't have the clear task -> work -> return result -> die model, rather it's more closely aligned to how I myself multitask, maintaining several long-lived threads of execution, and having to intentionally "forget" about the other ones when it's time to lock in and do real work on the most important one.
I have a few variations of pruning (trimming all long tool outputs, agent-directed trimming, etc) in my aichat collection of session tools:
https://pchalasani.github.io/claude-code-tools/tools/aichat/...
OMP has this, it's called "/shake" where bloat from tool calls and such is turned into an artifact ref instead of remaining in the context.
There are other pruning or compaction strategies you can configure too.
You can also use "/btw" to have side tangent conversations with the current context but with the benefit that it will not bloat your current context after the tangent is done.
I wish other harnesses were prioritising this sort of flexibility with context management, but they're likely more incentivised on optimising their cache rates
I do this in my own harness, where a context only exports certain messages when you step out of it — eg, my request for an essay and its final output, while dropping everything from the first draft through various intermediary edits.
That naturally trims a lot of context while removing mistakes from the context to prevent poisoning (ie, every draft but the final negatively contributes in some way — that’s why we edited them).
Potentially: remove thinking blocks, and keep the rest. At least this would ensure that the entire context of the conversation is still there, and anything said isn't lost.
Having a second model also iterate the resulting messages and remove low-value tool calls could also be interesting. Especially failed calls which add no value.
Wouldn’t that invalidate kv cache and you wouldn’t benefit from much cheaper cache read?
Oh my pi already prunes tool calls on compaction (not sure if normal pi also does this)
In oh-my-pi there is the /shake command that gets rids of tool results and some other things
check out this:
https://www.morphllm.com/products/compact
you can wire it into pi compaction pretty easily
I wonder if you could prune the kv cache as well
when you look at the compaction prompt: in a sense it is doing that pruning but the llm decides what to prune
Compaction is painful if you run just one local LLM, the best way to avoid it is to keep context as small as possible.
One trick I find useful is to have one model with two KV caches running and while first cache has produced tokens, second cache immediately summarizes them during input tokens are being generated (tools time), then harness switches to the second KV cache which takes newly produced input tokens while KV in first cache is getting replaced with compacted summary tokens. This is a kind of ping pong, so we trade more space for less time. Still experimenting but it looks it works, and nice bonus it improves GPU utilization. Btw I have my own harness and model serving code, but it can be easily implemented in any other harness and model server.
I tend to ask a line of questions to the LLM as first step to a point the context is enough for me. Then i ask for the work i need but then get back to previous context using /tree.
It's like asking the questions needed to get the answer you need from the LLM, can be either an answer for a plan/todo or a task to code/change something.
In my experience, the best approach to compaction is to never get to the point where you need compaction and to generally stay below about 30% context window utilization. Even for long agentic workflows this can be accomplished for quite a while, much longer than most people might think.
Here's what I do for each of my sessions:
1. For asides, off-topic work, or repetitive work that has already been done in the session, branch backwards (with /tree) and summarize.
2. If I've exceeded 30% or the 'price-doubling' multi-tier pricing, prune (my custom extension).
3. If I've already pruned and I'm still close to 30%, 'prune all' (more extensive prune).
Definition:
'/prune': Removes ~50% context on a fresh session (not previously pruned)
- Keeps: User messages, normal assistant prose, commands/status markers, extension receipts, model settings, and a plain-text receipt for each tool call.
- Removes: Thinking, signatures, actual tool calls/results, tool output, images, compaction summaries, and other extensions’ state.
- Keeps: User messages, normal assistant prose and conclusions, commands/status markers, extension receipts, and model settings.
- Removes: Thinking, signatures, all tool calls/results and output, images, compaction summaries, other extensions’ state, and any tool-activity receipts created by /prune.
Using these I can keep a session going for weeks (or longer), even with extensive use and almost all the important context is preserved while dumping the less important context. Neither command requires an LLM summarization so they execute quickly.
I am surprised at 'removes actual tool calls/results, tool output'. Your approach with /prune seems to be 'keep the WHAT, remove the HOW (we got here)'. I would have thought that the HOW contains some useful signal.
What is that 30% number based on? Surely that's a model specific limit, and is based more on the absolute token length, not percentage, right? I'm not sure it makes sense for e.g. Opus 0.2M and Opus 1M to both degrade at 30% of their respective context lengths.
People nowdays have no shame and completely forget the art of plug.
This is terrible. Models have been RLed on looking at the previous tool call chain, and reasoning. No chance this does not reduce performance. The point of compaction is that it also includes useful signal from the tool outputs itself so agent does not repeat it afterwards
I think the way prompt caching works really discourages more creative compaction techniques. Like perhaps some kind of heuristic progressive compaction that replaces tool results and thinking traces after use with pointers could potentially keep the model smart for much longer, but that'd mean breaking cache every turn, and possibly even within a turn, seriously driving up cost.
I don't like any of current solutions when it comes to compaction. I'd love to have a way to say what exactly should be summarized, because most of the time I just need to compact some noisy MCP tool calls, test runs and things like that. Just let me pick what should be summarized and keep the rest as is.
In juggler (https://news.ycombinator.com/item?id=48883305) I spent a lot of effort on some compaction tricks that are elegant, but in reality find myself not really using them..
Because juggler has a nested-thread architecture, I had a great ah-ha! moment when I realised that with sub-threads, compaction becomes almost free: you just take the list of items in the conversation, move them all into a new sub-thread, then allow that sub-thread to summarise itself (this summary is what the parent thread sees). Hey-presto, your parent thread is now compacted, and the sub-thread contains all the old messages, so you can browse them if you need to, or just undo the whole operation if you change you mind. This struck me as super-elegant, and I was chuffed that it works.
However.. What I find myself actually doing now is just using the "new conversation" tool to hand over to a new conversation - I just tell the LLM "create a new conversation, and tell it everything it needs to continue this task". It gets you to the same place, but generally I'm not interested in the history, and having a fresh conversation with a new name usually feels neater.
OMP changed the default compaction to images! Kinda nuts to read about. Saves the generation cost of the traditional compaction step and writes the context as tiny text to an image, if I was following correctly.
Great thread, I was just thinking about compaction. My current line of thought is that compaction/pruning/ctx management in general should be something ongoing and maybe recursive. For example:
User:'How is auth implemented?' -> [thinking] [codebase exploration with [thinking] in between, 10 file reads, 3 of which were "wrong"] [thinking] -> agent_response
This little exchange contains a WHAT (how auth actually is implemented) and a HOW (where that info is and how to retrieve it). Maybe this question was part of a larger task. I think that whole exchange could be summarised before it enters context, kind of like what happens with subagents. The main thread would then consist mostly of [summaries]. Eventually the context will fill up anyway and we would summarise those summaries again. Alternatively one could maintain a [master_summary], kind of like an internal state. So new [summaries] get integrated directly and the [master_summary] gets updated.
I implemented a few additional strategies in https://github.com/rcarmo/piclaw/tree/main/runtime/src/exten... - including Codex-native server-side compaction. They all have slightly different trade-offs, but I run very long sessions quite successfully
Was expecting the article to go more in-depth.
Say, what happens when chain of summaries grows so long, that it still overflows context window. Is summarization runned over the summaries in the context window?
The advantage of running local stack is that you can do the compaction at the time of inference, i.e. some tool call runs out of context, you can just pause inference, purge/replace old tool calls with their summaries or just logs by operating directly over tokens on a GPU, rebuilding KV cache (one time prefill hit) and resuming the inference, easily being able to e.g. read 1000 markdowns, each 50k long, in a single LLM call. That's not possible with current agentic harnesses using LLM calls.
Compaction has been a pretty painful part of local llm usage. Scrapping the current context and parsing almosy 128k of context then generating something like 5-10k tokens - that can take quite a while when you’re working with 10t/s-45t/s (depending on the model).
I pretty much just start a new session whenever i fill the context.
My summarization creation functions over batches of 50 messages, and I don’t often lose important context any more. The loss comes from trying to stick a whole conversation in a single compaction request (at least in my case)
How context management works in Swival https://swival.dev/pages/context-management.html
I don't like that it throws away the whole KV cache when compacting. It costs a cache miss of the whole conversation length, and that's a waste of time and money.
LLMs are perfectly capable of summarising the conversation without a new system prompt.
I want to like Pi but compaction is why I had to go back to Open Code…
My problem is that when in a loop and it’s calling tools, it won’t check how close it is to the compacting limit until the whole loop returns to you for the next prompt. And so if you have a run that could go for hours, it’s a gamble if you’ll OOM or an interrupting compaction breaks context and stops the loop without continuing.
There’s a few extensions that all try to solve this problem, but I’ve found none actually work :(
Does anyone find compaction useful? I have always been a big believer of many small chats instead of large ones. There is certainly a time and a place where maybe I need to load in a lot to initial context BUT I often find that quality goes does as we near context limits and of course costs go massively up. I would rather repeat small notarized chunks than have a massive context window.
I expect Pi is mostly used with OpenAI plans, and OpenAI has a dedicated compaction endpoint you should probably be using with their models instead of a compaction prompt.
There are some very interesting latent compaction approaches like this[1] for when you can control the whole inference stack. i.e in on-device and datacenter inference.
Is pi a drop in replacement for OpenCode / Claude Code? I cannot be bothered installing 50 million plugins.
Ampcode used a handoff feature for a while that I found genuinely useful [1] and then they removed it. Anecdotally, I felt it worked better than compaction.
the way compaction works is pretty simple and highly relies on another model to do it, the only part the user has control of is when to do the compaction, which actually means when doing large amount of work in one shot, it should be planned from the beginning to be separated into works that can be reviewed, afterwards the model can compact
Can someone recommend a Hermes alternative that is less token hungry? Pi did not work well for my use case.
TLDR: It keeps ~20k tokens of recent conversations, then hands the rest of the conversation to another model with a special system & user prompt. This then fills out a template with relevant information.
See: https://github.com/earendil-works/pi/blob/main/packages/codi...
Opencodes dynamic context pruning works by labeling tools and chat and the rest and the agent can collapse and expand summaries.
I get it into 1M+ routinely on local models with operations between 50k-85k