I apparently use Claude differently the most people who talk about using Claude on the internet.
I’ll typically have a bunch of short sessions over the course of a day. Anytime I start a task that isn’t going to very directly benefit from the existing context I start fresh.
I don’t find a lot of benefit in explaining the project overall to Claude — I’ve deleted a lot of that explanation from my Claude.md because it didn’t seem to impact much.
I typically start a task by pointing it to 1-2 files and giving it some explanation of what I want done, and it
figures it out.
Basically never hit context window limits or compactions, and can’t remember the last time I hit a 5 hour or a weekly limit.
I do something similar. I've recently started having a few "starting point" files to re-explain common context (less than thirty lines per markdown file, usually) that I can point the agent at at the start of a new session, each tightly scoped to a certain domain and/or task type. That's been nice to avoid repeating myself, without the side-tracking or over-aggressive biasing-towards-previous-conversations that I've seen happen if I use long sessions or let it try to decide on its own what to pull in from larger files or trees of files. Sometimes I'll tell it to update that file based on new info from a current task, but I keep tight control over what gets pulled into task start context.
They aren't really "explaining the project" either, but more module- or task-specific preferences, hand reference pointers, or other things like "there are mixed examples of how to do certain things in this project, prefer X to Y." I use a write-everything-twice approach. After I find myself having to correct an implementation because it didn't figure out one of these things on its own from the existing code, I'll add an entry. That also avoids bloating things with "I think this is relevant" compared to "I have noticed that this is necessary."
I keep doing this because it lets me experiment with different approaches to problems without risk of it fixating on things from a previous abandoned attempt, and particularly because sometimes I'm wrong and I haven't found the agent harnesses particularly reliable at taking my word for it from a POV of "yes I know I said we need xyz earlier, but let's please entirely forget about that."
This is my usage pattern and I agree it works really well. I start almost every conversation by asking Claude to read, not write. Then once it's explored a particular slice I let it rip.
This takes a couple minutes (and I suppose I'm spending tokens each time), but sessions rarely reach compaction length and I like that I'm not trying to keep a whole separate pile of docs in sync.
I keep a plan file that records what I’m doing, how I’m doing it, and what I’ve done so far - every time I sit down to have another session, I feed Claude the plan file first, then tell it to begin on the next unchecked todo. Every time I run into something new, I tell it to add it to the todos in the plan file.
It basically takes care of itself, or at least as close as it can.
Claude Code has a big system prompt, most of which isn't necessary for the more recent models. (Codex too.)
I've been running Claude and GPT in my own agent harness. The main difference I notice is that tasks take about 7x longer to complete if they're run in the official Claude or Codex harness (and cost me 7x more).
You would think this would lead to increased correctness, but that doesn't seem to be the case. Today I tested both side by side. They both resulted in data loss. (I had a backup obviously.)
GPT running in the official harness did a bunch of extra tests and double checking, and ended up with the same result regardless (it permanently deleted a bunch of documentation).
All else being equal, I like getting my data loss 7x faster and cheaper ;)
I work similar, but still have architecture mds for a few selected cross cutting features. As useful for human readers as for AI.
Normally I do it exactly as you say, point at a few files, but if I know these features are involved I point at the corresponding mds instead. Its a shortcut for me to type less.
Honestly sounds like you’re not doing anything difficult. If I’m doing easy tasks it’s fine but if you need to do a major architectural project that spans 3 codebases and 2 clouds you’re gonna have a hard time without substantial context/memory management.
I work in a completely different manner. I have a chief of staff agent, which is one Claude code instance that orchestrates work across all my projects simultaneously with sub agents. In this way the agent helps me context switch and drive work towards everything I’m working on. I only use 1 session, I compact when necessary with todos and on file system files to track wip
It's very simple. I have told Claude that it's my chief of staff and that it must delegate all tasks to subagents. That as my COS it must show initiative and only come to me for vision/decision making. I have hooks further enforcing this. It works really well because I am usually working on like 12 things at once. Feels like playing simultaneous chess. The COS agent helps me context switch/orchestrate everything. Subagents have their own context window so the COS context does not get polluted/filled up with low level subtask details. COS can effectively prompt the subagents. My main session does grow and I compact it from time to time. But since all the WIP is already captured in external state, compaction of the main session does not degrade performance.
I might be missing out on something but I never had to explain my project. Just give it a task, or if you really want to, type it quickly, then you are good to go.
I can’t imagine this being worth optimizing. The issue is never that Claude can’t figure out what the projects is about…
Am I missing something or does this project not solve a problem most regular people have?
There are many other posts here which agree with you. Filling context with what you think the model needs adds nothing and possibly just inflates context which is harmful.
A good method seems to be only make a skill or memory when the LLM gets something wrong, or if you actually observe it's always doing the same step and you can get the model to the same place with less tokens.
I’ve basically never edited a skill or memory myself. I make the LLM do it as part of the /handoff skill before I clear a session. That also includes pruning existing skills/memories and resolving any drift.
It's funny because with so many different implementations of /handoff, I wonder if anyone has benchmarked handoff-and-resume to figure out what the best performance implementation looks like.
Mine are project-specific, which is a bit annoying since I’d like it to be global but there are some project-specific additions. Maybe I’ll (ask Claude to) refactor that to be more composable.
It should be a first class feature of the harness, tbh. It kind of is with the /compact [focus] parameter but this is coarse and leaves no record. I find keeping the handoff files in the repo to be useful for historical context and later debugging.
> Filling context with what you think the model needs adds nothing and possibly just inflates context which is harmful.
The solution that I've developed is, let the agent figure things out efficiently, without inflating the context. I have what I call a smart repo that better explains this at
What I've finally come to understand is that there is a large amount of people who are now able to write and use software through claude and coding agents. Those people have different needs than more traditional software engineers who have more knowledge because even best llms often need steering, correction, and refactoring suggestions when iterating on code and it's fine to let it lose context because exactly like you said, you tell it to read file and then have to regurgitate the understanding so you can correct or validate it before continuing.
For those where the code is almost entirely a black box and cannot easily recover when something goes wrong. They are much more keen on this context management and planning because recovering from derailments is much harder (and takes longer) because its often a conversation with llm to try to recover to where they were before.
There are bunch of tools to manage context or fix what Claude does wrong. They may be popular because non software engineers want to improve their workflows, like you mentioned.
But are they really working instead of making it worse? Are there any tests or real case studies done by users not tool's author? From my experience, removing from context works more often then adding.
Depending on the scale of the project and the complexity of the specific thing you need to work on, it's advantageous to bring specific context into the session instead of hoping the model will connect the right dots.
I specifically have instructions for claude explaining the purpose of the project in pretty much all repos. Just a simple PROJECT.md is enough - and referenced from AGENTS.md
There I usually lay out stuff like "this is a personal greenfield project" and "don't bother with multi-user support" etc. Or Claude will default to creating something WEBSCALE for a simple tool that won't run outside of my local LAN-only Proxmox setup. And that'll also skip massive database migration support for a project that's 3 days old - the agent doesn't know that. I'm just dropping it on the project after a full memory wipe.
>I might be missing out on something but I never had to explain my project. Just give it a task, or if you really want to, type it quickly, then you are good to go.
This means its changes will either be out of alignment with the overall project and its "style" and goals, or it waste tokens re-getting to know the basics about the project each time.
Hm, I just keep a folder called something like `status_docs/` in any project I work on and I create a new file in that folder any day I'm working on a project that's dated (e.g. `project/status_docs/2026_06_21_status.md`). It's basically a project diary that both me and the LLM can reference.
I have the LLM at some point in the day while working on the project create that file with all the relevant context. And then I'll periodically have it update that file (often before I compact the context window, or before I switch to a new task). And then I just have the LLM update it whenever I'm done working on the project for the day.
Then no matter what, if I come back that project again a day later, a week later, a month later, whatever – I just literally point a fresh session at the most recent status doc to help both me and the LLM orient ourselves to the work at hand. What's really nice too is having it reference the status docs from previous days to help orient it for creating the new status doc for the current day.
I've been doing this informally for probably over a year now, and have started formalizing it so I do it with every project. It's been a big help to me personally given all the context switching between projects I've been doing more and more since using AI coding tools.
Somewhat related, I built ccrider [1], which indexes the session transcripts that agents already write to disk (Claude Code, Codex, Copilot) into a local SQLite FTS database. It has a TUI, CLI, and an MCP server so the agent can search past sessions itself. Same local-first idea, based the raw logs rather than a maintained digest. Should work nicely with recall as well.
I use Deepseek and just as it to generate a state.md file with a summary of the project every time I've reached a goal or milestone. I then take a few minutes to edit this and add in or take out details. Between token pricing and generous cache discounts This has proved very efficient so far, I reset every few hours of work and bypass the muddle of having too many priorities or over-extrapolation from un-nuanced instructions I gave at an earlier stage.
I do think that this project is interesting in several ways - prioritizing privacy, minimizing spend, and using objective semantic markers to sift and consolidate the key takeaways from long sessions. I'd like to try it on my cline project history. But while it would make a great recording of project history, I wonder if a lot of it doesn't end up detailing blind alleys the project went down and had to back out of.
Generally when this happens I feel that it's due to vague specification on my part, or avoiding architectural decisions I didn't want to deal with and implicitly inviting the model to implement a lowest-common-denominator solution.
> But while it would make a great recording of project history, I wonder if a lot of it doesn't end up detailing blind alleys the project went down and had to back out of
Yes, I've run into this as well. I had the agent document the changes it would make along the way, but it would want to keep making note of things that we switched away from. Eg, "In phase3 we used mysql instead of sqlite". On some projects I let the agwnt do all the documentation and plan.md and state.md files, and every so often go through manually and delete some of the cruft that's no longer needed.
The hard part with project memory isn’t saving more stuff, it’s deciding what not to trust later. Stale plans and failed debugging guesses can quietly poison an agent pretty fast.
I think the majority here have stated the same... That CLAUDE.md or AGENTS.md effectively do this. Either that or the readme.
The only tip I can give is that your skill that builds or wraps up work. You should have it update those files if anything has changed.
Claude/Agents files shouldn't be bloated, but should imho act as a basic amount of context on the project so your agent and skills can pick up and go, with even the most basic initial prompt.
> The only tip I can give is that your skill that builds or wraps up work. You should have it update those files if anything has changed.
Depending on the scope of work you’re doing, it might be better to have this removed from the context of the work that was done.
I keep a “Last Updated Hash” in my md and every so often will have the LLM pull a diff from that hash to the current head, then determine what doesn’t match.
Happy user of recall here. I
rarely need it as I try to keep conversations small and files-focused. But when I do need it, it brings a lot of value. Sometimes there are conversations where I failed to capture some interesting things. Recall is also very helpful to me to audit my system like when I start to suspect some inefficiencies around some tools (skills, mcps, clis). Recall was efficient to retrieve "tranversal context" required for such audit.
I have a documentation vault in my repo, organised using Obsidian (bases, wikilinks, frontmatter, etc), and accessible using obsidian-cli (and related Claude skills, thanks kepano). I started the repo agreeing with Claude the structure and front matter of documents and how to edit and read, all stored in a markdown file in the vault, and a specific instruction in the CLAUDE.md file on when and how to access it. Any updates require consistency sweeps. Any decisions made and agreed during implementation of new features get added to the vault.
It's been amazing how I just don't need to explain the project anymore. An empty context and a few sentences on what I'd like to spec next and the LLM finds what it needs in the vault.
Yes I've found this to be the case as well. I've also found that it's useful to split these documents into three broad distinguished classes: goals, design, and idioms.
The goal docs provide directionality - helping the agent generally make consistent design decisions. Scoping constraints and stuff are useful to put here, but also feature goals a general idea of where the project is heading long-term (even if none of the items there are on the implementation roadmap anytime soon). It keeps more of the sessions aligned with each other.
The design docs specify the state of the project as it is, and are kept in sync during implementation sessions, by instructing the agents to treat their updates as part of the implementation plan for any work.
The idioms docs keep track of incidental decisions that don't relate to long term goals per se, but things like code style detail, specific project-related investigative techniques, code organization rules, build process guides, etc.
It's a single anecdote but I found that overall the work encountered fewer low-level design mismatches where one session doing work on one thing would make a design decision that didn't really mesh well with another session doing work on another thing. Overall hygiene took less work to maintain.
There are likely a variety of superior ways of organizing things, but at the very least it seems there's a ton of value to be squeezed from just organizing your project meta-information in certain ways. Definitely worth spending some time experimenting with.
I have had a computer completely get fried when working on something. I did lose the project memory with Claude. Even when everything was in GitHub and was working in OneDrive it still took some time to get Claude up to speed on a second PC. Could see this being helpful.
anecdotal but in my last from-scratch project claude code seemed to pick up on a lot of nuance on its own — conventions, prior decisions, stuff i didn't explicitly restate — without me feeding it back in. makes me wonder how much of this gap is already closing on agent side vs needing a dedicated tool like this
My employer is counting token usage, so explaining my project between tokens isn’t necessarily a bad thing. I am clearly a more productive engineer because of it \end{sarcasm}
I never have to because I use a ticketing system the model goes through in addition to a CLAUDE.md file with a summary, including vision, goals, non-goals etc
Any tricks to get Claude to actually use the CLAUDE.md consistently? Many times now its completely ignored it, despite being short, concise + generated by Claude itself, and I see bug reports about this that are over a year old
If you JUST updated it, have Claude re-read it, this was the biggest thing for me. If you just created it, same thing. Even though Claude created the file, it has not read it in whatever special way Claude will read that file, which is silly, but once you realize this, and have Claude read it, it makes a difference. It will always load it from a new session, but if it changes or doesn't exist when you start a session it will be stale or empty.
Check out your session logs and review what is actually in the context window. I’m willing to bet that your CLAUDE.md is sitting close to the middle of everything in there. The current gen of frontier models tends to heavily weight the start and end of the current context so heavily that anything partway through may just be ignored.
What behavioural things are you noticing that Claude does not pickup from the CLAUDE.md. I am working on a `pi-brains` extension for the Pi agent. It is designed to inject rules into write and edits tool calls for matching files.
I am curious if the behaviour you want is outside of writing and editing files.
Love to see people creating such projects, i'm truly passionate about Coding Agent Memory myself.
Have you thought on making it collaborative memory across your teammates working on the repo?
how does your framework handle stale memory, when someone changes the code that breaks the memory store?
Thank you for the encouragement. Currently there is no way to handle stale memory and I didnt think about the teammates angle...good point I will need to think about it and see if I can build something around it.
Working on a project connected through multiple repos I have a project level MD file to explain each repo.
Then I let Claude create a hook it self that at the end of each conversation it starts a hook to add important stuff automatically to memory if relevant. That’s it…
CLAUDE.md is already a good system for context window management for all the same reasons that version control management of code is good.
And keeping a local copy of everything you ever told Claude in your context window is bad for the same reasons keeping a local copy of your code called My_Code_v3_final.zip is bad.
I was tired of seeing "--resume"^W loose so much context.
[edit: not --resume, I meant a new session using project memory]
I had the idea of an oracle/apprentice.
The idea was that the new session would learn from the old session, like a tutor.
I asked Claude to code a program, so that the old session would launch in loop, being the "oracle", and the new session would use it to connect to the oracle, to ask it questions.
So when I'm doing that, I'm seeing the two Claude sessions discussing between them. It's fun seeing the "apprentice" asking follow-ups questions.
The mechanism (on a Linux machine) is relatively simple : I asked Claude to code a Go tool, to use an abstract socket (\0claude-handoff-<projectname>), and specified that whichever is the first to successfully listen (no EADDRINUSE) becomes the "listen()er/accept()er" and that the second becomes the "connect()er".
So that establishing the socket in whichever order is independent of which of them is the oracle/listener.
I've put the mechanism in a global Claude rules. In the oracle when I'm a 98% usage of the 1M context, I just have to type "handoff <projectname> oracle", and to start a new session with "handoff <projectname> client".
And the "oracle" will loop on the tool (with a subagent, waiting indefinitely), the tool exits with a question on output, and re-call (with a subagent "handoff <projectname> answer") to give back the answer (which automatically waits for the next question).
And since the oracle is doing the call to the handoff tool in a subagent, when you see it answering, you can also type something along the lines of "hey please also precise to the apprentice <some specific information>".
The "transmission of knowledge which matters" is so much efficient, that ~2% of remaining context (20k tokens) is enough to transmis WAY MORE USEFUL information that any memory saving which would miss important informations.
It's not unexpected. It's like real-life. You may have an human put all informations that you want in some documentation, nothing can replace a phone line from the new human to the previous human for specific follow-up questions.
Tho one of the mattering rule specificity is to precise that the oracle should always include in this response the level of confidence in the answer, like if it is certainty/guess/hypothesis.
It's fun seeing the two Claudes discussing like two colleagues. I guess you could also ask Claude to code the tool to instead connect on a localhost IRC server.
I think that if you want this tool, you just have to c/c my text into a new Claude session and to tell it "I want this too, please code it, please also setup the global rule".
Another day another "memory" system for a tool that cannot ever have memory. LLMs have context, and the more you fill that context with unrelated junk, the worse they perform.
I apparently use Claude differently the most people who talk about using Claude on the internet.
I’ll typically have a bunch of short sessions over the course of a day. Anytime I start a task that isn’t going to very directly benefit from the existing context I start fresh.
I don’t find a lot of benefit in explaining the project overall to Claude — I’ve deleted a lot of that explanation from my Claude.md because it didn’t seem to impact much.
I typically start a task by pointing it to 1-2 files and giving it some explanation of what I want done, and it figures it out.
Basically never hit context window limits or compactions, and can’t remember the last time I hit a 5 hour or a weekly limit.
I'm the same. Every time I see people talking about explaining the project every session I'm puzzled. I've just never had to do that.
I do something similar. I've recently started having a few "starting point" files to re-explain common context (less than thirty lines per markdown file, usually) that I can point the agent at at the start of a new session, each tightly scoped to a certain domain and/or task type. That's been nice to avoid repeating myself, without the side-tracking or over-aggressive biasing-towards-previous-conversations that I've seen happen if I use long sessions or let it try to decide on its own what to pull in from larger files or trees of files. Sometimes I'll tell it to update that file based on new info from a current task, but I keep tight control over what gets pulled into task start context.
They aren't really "explaining the project" either, but more module- or task-specific preferences, hand reference pointers, or other things like "there are mixed examples of how to do certain things in this project, prefer X to Y." I use a write-everything-twice approach. After I find myself having to correct an implementation because it didn't figure out one of these things on its own from the existing code, I'll add an entry. That also avoids bloating things with "I think this is relevant" compared to "I have noticed that this is necessary."
I keep doing this because it lets me experiment with different approaches to problems without risk of it fixating on things from a previous abandoned attempt, and particularly because sometimes I'm wrong and I haven't found the agent harnesses particularly reliable at taking my word for it from a POV of "yes I know I said we need xyz earlier, but let's please entirely forget about that."
This is my usage pattern and I agree it works really well. I start almost every conversation by asking Claude to read, not write. Then once it's explored a particular slice I let it rip.
This takes a couple minutes (and I suppose I'm spending tokens each time), but sessions rarely reach compaction length and I like that I'm not trying to keep a whole separate pile of docs in sync.
Claude seems to read when it needs to when I ask it to do something. Did you have a different experience?
I keep a plan file that records what I’m doing, how I’m doing it, and what I’ve done so far - every time I sit down to have another session, I feed Claude the plan file first, then tell it to begin on the next unchecked todo. Every time I run into something new, I tell it to add it to the todos in the plan file.
It basically takes care of itself, or at least as close as it can.
In some sessions I find claude to be incredibly dumb - other times we are on the same wavelength and everything flows.
I guess I need to do some claude.md work or find other ways to prime the session so i get the good personality and not the evil twin.
Claude Code has a big system prompt, most of which isn't necessary for the more recent models. (Codex too.)
I've been running Claude and GPT in my own agent harness. The main difference I notice is that tasks take about 7x longer to complete if they're run in the official Claude or Codex harness (and cost me 7x more).
You would think this would lead to increased correctness, but that doesn't seem to be the case. Today I tested both side by side. They both resulted in data loss. (I had a backup obviously.)
GPT running in the official harness did a bunch of extra tests and double checking, and ended up with the same result regardless (it permanently deleted a bunch of documentation).
All else being equal, I like getting my data loss 7x faster and cheaper ;)
So you are using the API directly in your own harness without the subscription?
Yeah, I made a simple agent based on this tutorial:
https://minimal-agent.com/
It's slightly bigger now, but here's a ~50 line version for reference. I added the missing outer while-loop, so it takes user input etc.
https://gist.github.com/a-n-d-a-i/bd50aaa4bdb15f9a4cc8176ee3...
I mostly use it with GLM via their coding plan, I got a year for like $20 when it was on sale. But I also hooked it up to Sonnet, Opus, GPT, etc.
I work similar, but still have architecture mds for a few selected cross cutting features. As useful for human readers as for AI.
Normally I do it exactly as you say, point at a few files, but if I know these features are involved I point at the corresponding mds instead. Its a shortcut for me to type less.
Honestly sounds like you’re not doing anything difficult. If I’m doing easy tasks it’s fine but if you need to do a major architectural project that spans 3 codebases and 2 clouds you’re gonna have a hard time without substantial context/memory management.
But that's an anomaly. I'm pretty sure you're not doing major architectural changes over 3 codebases and 2 clouds daily?
you'd be surprised.
I work in a completely different manner. I have a chief of staff agent, which is one Claude code instance that orchestrates work across all my projects simultaneously with sub agents. In this way the agent helps me context switch and drive work towards everything I’m working on. I only use 1 session, I compact when necessary with todos and on file system files to track wip
Mind sharing details of your setup?
It's very simple. I have told Claude that it's my chief of staff and that it must delegate all tasks to subagents. That as my COS it must show initiative and only come to me for vision/decision making. I have hooks further enforcing this. It works really well because I am usually working on like 12 things at once. Feels like playing simultaneous chess. The COS agent helps me context switch/orchestrate everything. Subagents have their own context window so the COS context does not get polluted/filled up with low level subtask details. COS can effectively prompt the subagents. My main session does grow and I compact it from time to time. But since all the WIP is already captured in external state, compaction of the main session does not degrade performance.
I have used this same pattern in my own harness and it works well there too. https://github.com/computerex/z
I hooked an instance of my harness up to telegram and now I talk to it from everywhere it and dispatches work out to subagents.
I might be missing out on something but I never had to explain my project. Just give it a task, or if you really want to, type it quickly, then you are good to go.
I can’t imagine this being worth optimizing. The issue is never that Claude can’t figure out what the projects is about…
Am I missing something or does this project not solve a problem most regular people have?
There are many other posts here which agree with you. Filling context with what you think the model needs adds nothing and possibly just inflates context which is harmful.
A good method seems to be only make a skill or memory when the LLM gets something wrong, or if you actually observe it's always doing the same step and you can get the model to the same place with less tokens.
I’ve basically never edited a skill or memory myself. I make the LLM do it as part of the /handoff skill before I clear a session. That also includes pruning existing skills/memories and resolving any drift.
Even the /handoff skill was written by the model…
It's funny because with so many different implementations of /handoff, I wonder if anyone has benchmarked handoff-and-resume to figure out what the best performance implementation looks like.
I also imagine that varies by model.
Mine are project-specific, which is a bit annoying since I’d like it to be global but there are some project-specific additions. Maybe I’ll (ask Claude to) refactor that to be more composable.
It should be a first class feature of the harness, tbh. It kind of is with the /compact [focus] parameter but this is coarse and leaves no record. I find keeping the handoff files in the repo to be useful for historical context and later debugging.
> Filling context with what you think the model needs adds nothing and possibly just inflates context which is harmful.
The solution that I've developed is, let the agent figure things out efficiently, without inflating the context. I have what I call a smart repo that better explains this at
https://github.com/gitsense/smart-ripgrep
The basic idea is, when the agent does a ripgrep it gets back files + matching lines + context.
What I've finally come to understand is that there is a large amount of people who are now able to write and use software through claude and coding agents. Those people have different needs than more traditional software engineers who have more knowledge because even best llms often need steering, correction, and refactoring suggestions when iterating on code and it's fine to let it lose context because exactly like you said, you tell it to read file and then have to regurgitate the understanding so you can correct or validate it before continuing.
For those where the code is almost entirely a black box and cannot easily recover when something goes wrong. They are much more keen on this context management and planning because recovering from derailments is much harder (and takes longer) because its often a conversation with llm to try to recover to where they were before.
There are bunch of tools to manage context or fix what Claude does wrong. They may be popular because non software engineers want to improve their workflows, like you mentioned.
But are they really working instead of making it worse? Are there any tests or real case studies done by users not tool's author? From my experience, removing from context works more often then adding.
Git commit?
Depending on the scale of the project and the complexity of the specific thing you need to work on, it's advantageous to bring specific context into the session instead of hoping the model will connect the right dots.
I specifically have instructions for claude explaining the purpose of the project in pretty much all repos. Just a simple PROJECT.md is enough - and referenced from AGENTS.md
There I usually lay out stuff like "this is a personal greenfield project" and "don't bother with multi-user support" etc. Or Claude will default to creating something WEBSCALE for a simple tool that won't run outside of my local LAN-only Proxmox setup. And that'll also skip massive database migration support for a project that's 3 days old - the agent doesn't know that. I'm just dropping it on the project after a full memory wipe.
>I might be missing out on something but I never had to explain my project. Just give it a task, or if you really want to, type it quickly, then you are good to go.
This means its changes will either be out of alignment with the overall project and its "style" and goals, or it waste tokens re-getting to know the basics about the project each time.
No third case.
I guess that depends on the kind of project, how common the intent is, how self contained, etc.
Sometimes its good to start fresh. LLMs need large context restart's sometimes so they can better identify holes that they become blind to.
Back in the human age of coding, I felt the same way sometimes
This was always one of the reasons to hire interns
what a depressing statement.
Hm, I just keep a folder called something like `status_docs/` in any project I work on and I create a new file in that folder any day I'm working on a project that's dated (e.g. `project/status_docs/2026_06_21_status.md`). It's basically a project diary that both me and the LLM can reference.
I have the LLM at some point in the day while working on the project create that file with all the relevant context. And then I'll periodically have it update that file (often before I compact the context window, or before I switch to a new task). And then I just have the LLM update it whenever I'm done working on the project for the day.
Then no matter what, if I come back that project again a day later, a week later, a month later, whatever – I just literally point a fresh session at the most recent status doc to help both me and the LLM orient ourselves to the work at hand. What's really nice too is having it reference the status docs from previous days to help orient it for creating the new status doc for the current day.
I've been doing this informally for probably over a year now, and have started formalizing it so I do it with every project. It's been a big help to me personally given all the context switching between projects I've been doing more and more since using AI coding tools.
IntelliJ handles this for you. Basically it sends half your project to Claude even if you're asking some question about Star Wars.
> IntelliJ handles this for you. Basically it sends half your project to Claude
Not sure I’d call that “stopping wasting my tokens”.
Yeah but it could have sent the other half of the project too.
Somewhat related, I built ccrider [1], which indexes the session transcripts that agents already write to disk (Claude Code, Codex, Copilot) into a local SQLite FTS database. It has a TUI, CLI, and an MCP server so the agent can search past sessions itself. Same local-first idea, based the raw logs rather than a maintained digest. Should work nicely with recall as well.
[1] https://github.com/neilberkman/ccrider
I use Deepseek and just as it to generate a state.md file with a summary of the project every time I've reached a goal or milestone. I then take a few minutes to edit this and add in or take out details. Between token pricing and generous cache discounts This has proved very efficient so far, I reset every few hours of work and bypass the muddle of having too many priorities or over-extrapolation from un-nuanced instructions I gave at an earlier stage.
I do think that this project is interesting in several ways - prioritizing privacy, minimizing spend, and using objective semantic markers to sift and consolidate the key takeaways from long sessions. I'd like to try it on my cline project history. But while it would make a great recording of project history, I wonder if a lot of it doesn't end up detailing blind alleys the project went down and had to back out of.
Generally when this happens I feel that it's due to vague specification on my part, or avoiding architectural decisions I didn't want to deal with and implicitly inviting the model to implement a lowest-common-denominator solution.
> But while it would make a great recording of project history, I wonder if a lot of it doesn't end up detailing blind alleys the project went down and had to back out of
Yes, I've run into this as well. I had the agent document the changes it would make along the way, but it would want to keep making note of things that we switched away from. Eg, "In phase3 we used mysql instead of sqlite". On some projects I let the agwnt do all the documentation and plan.md and state.md files, and every so often go through manually and delete some of the cruft that's no longer needed.
The hard part with project memory isn’t saving more stuff, it’s deciding what not to trust later. Stale plans and failed debugging guesses can quietly poison an agent pretty fast.
"correctly invalidating your cache" rears it's ugly head once again.
How does this beat a Session specific README?
I think the majority here have stated the same... That CLAUDE.md or AGENTS.md effectively do this. Either that or the readme.
The only tip I can give is that your skill that builds or wraps up work. You should have it update those files if anything has changed.
Claude/Agents files shouldn't be bloated, but should imho act as a basic amount of context on the project so your agent and skills can pick up and go, with even the most basic initial prompt.
> The only tip I can give is that your skill that builds or wraps up work. You should have it update those files if anything has changed.
Depending on the scope of work you’re doing, it might be better to have this removed from the context of the work that was done.
I keep a “Last Updated Hash” in my md and every so often will have the LLM pull a diff from that hash to the current head, then determine what doesn’t match.
With all due respect, this github repo looks really like an AI-generated project.
The question I find myself asking very often these days: Is this better than asking claude to do the same things the plugin/repo does?
Doesn't Claude have memories like Codex?
It does.
Edit: oh no, so sorry, I am using another project named recall. But not this one. https://github.com/arjunkmrm/recall
Happy user of recall here. I rarely need it as I try to keep conversations small and files-focused. But when I do need it, it brings a lot of value. Sometimes there are conversations where I failed to capture some interesting things. Recall is also very helpful to me to audit my system like when I start to suspect some inefficiencies around some tools (skills, mcps, clis). Recall was efficient to retrieve "tranversal context" required for such audit.
I have a documentation vault in my repo, organised using Obsidian (bases, wikilinks, frontmatter, etc), and accessible using obsidian-cli (and related Claude skills, thanks kepano). I started the repo agreeing with Claude the structure and front matter of documents and how to edit and read, all stored in a markdown file in the vault, and a specific instruction in the CLAUDE.md file on when and how to access it. Any updates require consistency sweeps. Any decisions made and agreed during implementation of new features get added to the vault. It's been amazing how I just don't need to explain the project anymore. An empty context and a few sentences on what I'd like to spec next and the LLM finds what it needs in the vault.
Yes I've found this to be the case as well. I've also found that it's useful to split these documents into three broad distinguished classes: goals, design, and idioms.
The goal docs provide directionality - helping the agent generally make consistent design decisions. Scoping constraints and stuff are useful to put here, but also feature goals a general idea of where the project is heading long-term (even if none of the items there are on the implementation roadmap anytime soon). It keeps more of the sessions aligned with each other.
The design docs specify the state of the project as it is, and are kept in sync during implementation sessions, by instructing the agents to treat their updates as part of the implementation plan for any work.
The idioms docs keep track of incidental decisions that don't relate to long term goals per se, but things like code style detail, specific project-related investigative techniques, code organization rules, build process guides, etc.
It's a single anecdote but I found that overall the work encountered fewer low-level design mismatches where one session doing work on one thing would make a design decision that didn't really mesh well with another session doing work on another thing. Overall hygiene took less work to maintain.
There are likely a variety of superior ways of organizing things, but at the very least it seems there's a ton of value to be squeezed from just organizing your project meta-information in certain ways. Definitely worth spending some time experimenting with.
I have had a computer completely get fried when working on something. I did lose the project memory with Claude. Even when everything was in GitHub and was working in OneDrive it still took some time to get Claude up to speed on a second PC. Could see this being helpful.
anecdotal but in my last from-scratch project claude code seemed to pick up on a lot of nuance on its own — conventions, prior decisions, stuff i didn't explicitly restate — without me feeding it back in. makes me wonder how much of this gap is already closing on agent side vs needing a dedicated tool like this
How's it different to https://github.com/angelnicolasc/graymatter ?
My employer is counting token usage, so explaining my project between tokens isn’t necessarily a bad thing. I am clearly a more productive engineer because of it \end{sarcasm}
I never have to because I use a ticketing system the model goes through in addition to a CLAUDE.md file with a summary, including vision, goals, non-goals etc
If it goes through all of it all the time, then it fits the “wasting token” description.
Any tricks to get Claude to actually use the CLAUDE.md consistently? Many times now its completely ignored it, despite being short, concise + generated by Claude itself, and I see bug reports about this that are over a year old
If you JUST updated it, have Claude re-read it, this was the biggest thing for me. If you just created it, same thing. Even though Claude created the file, it has not read it in whatever special way Claude will read that file, which is silly, but once you realize this, and have Claude read it, it makes a difference. It will always load it from a new session, but if it changes or doesn't exist when you start a session it will be stale or empty.
Check out your session logs and review what is actually in the context window. I’m willing to bet that your CLAUDE.md is sitting close to the middle of everything in there. The current gen of frontier models tends to heavily weight the start and end of the current context so heavily that anything partway through may just be ignored.
What behavioural things are you noticing that Claude does not pickup from the CLAUDE.md. I am working on a `pi-brains` extension for the Pi agent. It is designed to inject rules into write and edits tool calls for matching files.
I am curious if the behaviour you want is outside of writing and editing files.
Anything in particular? Paca? I used forgejo for a while and it was OK.
Love to see people creating such projects, i'm truly passionate about Coding Agent Memory myself.
Have you thought on making it collaborative memory across your teammates working on the repo? how does your framework handle stale memory, when someone changes the code that breaks the memory store?
Thank you for the encouragement. Currently there is no way to handle stale memory and I didnt think about the teammates angle...good point I will need to think about it and see if I can build something around it.
Working on a project connected through multiple repos I have a project level MD file to explain each repo.
Then I let Claude create a hook it self that at the end of each conversation it starts a hook to add important stuff automatically to memory if relevant. That’s it…
CLAUDE.md is already a good system for context window management for all the same reasons that version control management of code is good.
And keeping a local copy of everything you ever told Claude in your context window is bad for the same reasons keeping a local copy of your code called My_Code_v3_final.zip is bad.
If I need context for a session then that is output from a previous session, otherwise I find any “memory” functionality cumbersome.
I saw /graphify recently which cuts down on exploration cost and seems more appealing (although I haven’t tried it yet)
Nothing wrong with a toy project.
Exciting to see hooks used for automation.
But if I may, the need to manually update the context is a huge hurdle.
Automation like this is limited unless no human has to remember it. So perhaps you can save context during the PreCompact and Stop hooks.
This whole project assumes that --resume replays full transcript, is that actually true? Is there no caching going on?
Are there any benchmarks/evals to back the claims? Or how do you know that it helps reducing waste?
They don’t. It’s all based on FOMO. Like superpowers, the more the better.
My advice: the best claude is the raw claude, with some custom tailored skills. That’s it, no plugins.
Why would you want a simple summarizer instead of frontier AI doing the summarization for you?
Bc frontier models are expensive and summarization is basic af
Compaction already drops nuanced and salient information. I wouldn’t want any additional amnesia from overly simplistic summarization.
Has someone tested this paired with graphify ?
I was tired of seeing "--resume"^W loose so much context. [edit: not --resume, I meant a new session using project memory]
I had the idea of an oracle/apprentice. The idea was that the new session would learn from the old session, like a tutor.
I asked Claude to code a program, so that the old session would launch in loop, being the "oracle", and the new session would use it to connect to the oracle, to ask it questions.
So when I'm doing that, I'm seeing the two Claude sessions discussing between them. It's fun seeing the "apprentice" asking follow-ups questions.
The mechanism (on a Linux machine) is relatively simple : I asked Claude to code a Go tool, to use an abstract socket (\0claude-handoff-<projectname>), and specified that whichever is the first to successfully listen (no EADDRINUSE) becomes the "listen()er/accept()er" and that the second becomes the "connect()er".
So that establishing the socket in whichever order is independent of which of them is the oracle/listener.
I've put the mechanism in a global Claude rules. In the oracle when I'm a 98% usage of the 1M context, I just have to type "handoff <projectname> oracle", and to start a new session with "handoff <projectname> client".
And the "oracle" will loop on the tool (with a subagent, waiting indefinitely), the tool exits with a question on output, and re-call (with a subagent "handoff <projectname> answer") to give back the answer (which automatically waits for the next question).
And since the oracle is doing the call to the handoff tool in a subagent, when you see it answering, you can also type something along the lines of "hey please also precise to the apprentice <some specific information>".
The "transmission of knowledge which matters" is so much efficient, that ~2% of remaining context (20k tokens) is enough to transmis WAY MORE USEFUL information that any memory saving which would miss important informations.
It's not unexpected. It's like real-life. You may have an human put all informations that you want in some documentation, nothing can replace a phone line from the new human to the previous human for specific follow-up questions.
Tho one of the mattering rule specificity is to precise that the oracle should always include in this response the level of confidence in the answer, like if it is certainty/guess/hypothesis.
It's fun seeing the two Claudes discussing like two colleagues. I guess you could also ask Claude to code the tool to instead connect on a localhost IRC server.
I think that if you want this tool, you just have to c/c my text into a new Claude session and to tell it "I want this too, please code it, please also setup the global rule".
Willing to try this but the author missed that Claude also has memory.md
The author appears to be Claude.
How is this different than just using a resume?
Anyone tried this?
Another day another "memory" system for a tool that cannot ever have memory. LLMs have context, and the more you fill that context with unrelated junk, the worse they perform.