Although I’d like this to work, I have not seen evidence that AIs extract good reusable knowledge from sessions without strong guidance.
I do R&D work in comp sci and usually write software and reports or papers in parallel. I have been using Opus in a controlled human-in-the-loop fashion to (significantly) speed up the work. While I’m at times surprised how high-level input steers output the right way, no high level ideas emerge from the AIs output. Many attempts to abstract from the concrete are wrong.
As a consequence, most writing is poor on content and form. It writes about the wrong things and crosses abstraction levels all the time. It cannot keep implementation details and conceptual leaps apart.
I don’t know what this means for the aim of this project to maintain important info for agents. Perhaps there are enough low-hanging fruits. That said, I’m skeptical that the resulting wiki is good documentation for human contributors.
Understanding what is relevant to be documented vs not is definitely tricky and something we've spent a lot of time on. We found being more prescriptive about what should be documented is helpful, we incorporated guidance from diataxis for it (https://diataxis.fr/)
Also the wiki is primarily for your agents and not for people. Although we're working to improve the prose quality and structure so it is well written for humans too.
Do you mean that specifically those models perform badly without guidance or do all models need specific instructions for technical writing?
We have observed that it is a loss for teamwork that the sparring with AI is entirely lost when only the technical conclusion is recorded. I write to recover that for my team and an AI can use it to its advantage as a side effect.
It would be cool to have good tool support for this process, but I’d bet on support, not delegation in that case.
The models perform badly without guidance but we've found the guidance doesn't need a human in the loop (at least for this task). You can distill what makes a good wiki, what information should go in etc. and the models are good enough to do it reliably. That's what took the most time (researching and studying existing wikis)
When you say support do you mean something where you can collaborate with the model to maintain the wiki?
Glad you liked it! Would love to chat more about what kind of production workflows you are looking to improve. You could grab whichever time is convenient for you: https://cal.com/team/almanac/demo
LoCoMo is not quite our use case. It tests conversational memory, while Almanac is more for coding agents trying to find their way around a codebase.
These results show output quality at the same token budget, not token savings yet. We’re working on coding agent evals that should be more representative.
Not in any way being negative about what you’ve done, but… Mac only. Is this just because of inbuilt knowledge about file system shape or some more important Apple exclusive tool/environmental quirk?
Although I’d like this to work, I have not seen evidence that AIs extract good reusable knowledge from sessions without strong guidance.
I do R&D work in comp sci and usually write software and reports or papers in parallel. I have been using Opus in a controlled human-in-the-loop fashion to (significantly) speed up the work. While I’m at times surprised how high-level input steers output the right way, no high level ideas emerge from the AIs output. Many attempts to abstract from the concrete are wrong.
As a consequence, most writing is poor on content and form. It writes about the wrong things and crosses abstraction levels all the time. It cannot keep implementation details and conceptual leaps apart.
I don’t know what this means for the aim of this project to maintain important info for agents. Perhaps there are enough low-hanging fruits. That said, I’m skeptical that the resulting wiki is good documentation for human contributors.
We've found Opus/Gpt are really bad at technical writing. We've found ways around this by having guidelines like https://www.openalmanac.org/ai-patterns-to-avoid.md, which have worked quite well for us.
Understanding what is relevant to be documented vs not is definitely tricky and something we've spent a lot of time on. We found being more prescriptive about what should be documented is helpful, we incorporated guidance from diataxis for it (https://diataxis.fr/)
Also the wiki is primarily for your agents and not for people. Although we're working to improve the prose quality and structure so it is well written for humans too.
Do you mean that specifically those models perform badly without guidance or do all models need specific instructions for technical writing?
We have observed that it is a loss for teamwork that the sparring with AI is entirely lost when only the technical conclusion is recorded. I write to recover that for my team and an AI can use it to its advantage as a side effect.
It would be cool to have good tool support for this process, but I’d bet on support, not delegation in that case.
The models perform badly without guidance but we've found the guidance doesn't need a human in the loop (at least for this task). You can distill what makes a good wiki, what information should go in etc. and the models are good enough to do it reliably. That's what took the most time (researching and studying existing wikis)
When you say support do you mean something where you can collaborate with the model to maintain the wiki?
Yes: the human provides the qualitative judgement and the agent provides the executive function.
This looks promising. Is there a version I can use to self improve my production workflows and not just the dev work?
Glad you liked it! Would love to chat more about what kind of production workflows you are looking to improve. You could grab whichever time is convenient for you: https://cal.com/team/almanac/demo
Any data on better outputs and/or token saving?
We’re working on evals now. We ran a small preliminary LoCoMo test with a 2k-token retrieval budget:
Almanac: 55.7% BM25: 51.8% Supermemory: 47.6% Mem0: 60.6%
LoCoMo is not quite our use case. It tests conversational memory, while Almanac is more for coding agents trying to find their way around a codebase.
These results show output quality at the same token budget, not token savings yet. We’re working on coding agent evals that should be more representative.
Cool thanks I will keep an eye on it!
Not in any way being negative about what you’ve done, but… Mac only. Is this just because of inbuilt knowledge about file system shape or some more important Apple exclusive tool/environmental quirk?
We’re working on Linux/Windows support right now. Mac was the easiest to setup for us since that’s the machine we code on. What OS do you use?