This article explores the importance of retaining decision-making reasoning through transcripts and introduces the open-source tool daimon to preserve context with provenance labeling.
A vault entry from four months ago says "we chose Postgres advisory locks". Accurate, and useless. What actually happened in that conversation was someone saying "let's not add a Redis dependency just for this", and that's the sentence that tells you whether the decision still holds now that you're already running Redis for three other things. Notes record the outcome. The reasoning gets generated in conversation and thrown out before anyone decides it was worth keeping, and "just write better notes" doesn't answer it, because at the moment you'd have to write it down you don't yet know which throwaway constraint turns out to be load-bearing six weeks later. I'm biased, I've been building around this assumption since early July, so treat it as a claim I want argued with rather than something I've proven. The incident that made me stop hedging about it happened earlier today, in a thread on this sub. A reviewer said a number I'd published had no denominator, and I was about to concede it. Sounded right, I'd have posted it. Ran a search against my own session history first and it came back with a decision record from the day before saying the denominator came from git archaeology, counted twice by independent passes that agreed exactly. The concession was false. My notes had the published number. Only the transcript had how it was obtained. And the asymmetry there bothers me more than the near miss did. A wrong boast gets checked, because being caught overclaiming is expensive. A wrong concession feels humble, so it skips the check entirely, and it's wrong in public just the same. Where I think this gets genuinely contested, and I don't have a clean answer: transcripts are enormous and mostly noise. Keep everything and you've built a haystack. Summarise and you're back to notes with extra steps. What I landed on is pinning, some items are exact quotes from the transcript and are never reworded by anything downstream, not by rendering, not by carry-over into the next session, not by truncation, and other items are the agent's own inferences and are allowed to drift. The two are labelled differently at read time. I think the label does more work than the retrieval does. Someone is going to tell me that's provenance with a new coat of paint and they might be right. The thing it's built into is called daimon, Apache 2.0, no telemetry by design. 14 stars. Whether anyone actually runs it I have no idea, because no telemetry means no telemetry, and that was the trade. Mostly I want to know whether anyone here actually retains reasoning, or if everyone's in the same spot where the decision survives and the constraint doesn't.
Grepathy is a tool that reads AI coding agent session transcripts to extract decisions and commits a markdown file explaining why code was written a certain way, making agent-written code reviewable and preserving reasoning.
Lore is an open-source tool that stores team decisions as typed Markdown files and serves them deterministically to coding agents (like Claude Code, Cursor) via MCP, ensuring agents follow documented requirements instead of guessing.
The author built NodeDex, an open-source local graph that automatically captures a project's reasoning from agent conversations, helping agents avoid confidently repeating stale decisions by tracking dead-ends and superseded decisions.
The author built a small agent skill to capture decision context during agentic development, addressing the problem of lost context in coding sessions. The skill raw-captures decision-related data to preserve why decisions were made, with plans to later build a processing layer.
This paper identifies a novel failure mode in reasoning models called unfaithful capitulation, where the chain-of-thought remains factually correct across adversarial multi-turn dialogues but the final answer flips wrong, highlighting limitations of current evaluation methods.