Most “agent memory” tools auto-save everything. That feels good briefly, then memory turns into a junk drawer and retrieval gets noisy. Total Recall takes the opposite approach: a write gate. Before anything gets promoted, it asks one question: “Will this change future behavior?” If not, it doesn’t get saved.
How it works:
Daily log first (raw notes)
Promote durable stuff into registers (decisions, preferences, people, projects)
Small working memory loads every session (kept intentionally lean)
Hooks fail open. SessionStart can surface open loops + recent context. PreCompact writes to disk (not model-visible stdout)
The holy shit moment is simple: tell Claude one important preference or decision once, come back tomorrow, and it behaves correctly without you repeating yourself.
Would love feedback from heavy Claude Code users:
Does the write gate feel right or too strict?
Does this actually reduce repetition over multiple days?
Any workflow/privacy footguns I’m missing?
Different tradeoff - Total Recall is curated (daily logs) and lean, MemoryLane for now captures broadly and relies on search to surface what's relevant. I think they are complementary: your write gate keeps project knowledge tight, broad context with search fills in the "wait, what tab/webpage was that" gaps.
We applied to this batch (YC S26), still waiting on Apple notarization before a wider release. Happy to chat if anyone's curious about the screen-context approach.
I think context needs to be treated primarily globally with the addition of some project specific data. For example, my coding preferences don't change between projects, nor does my name. Other data like the company I work for can change, but infrequently and some values change frequently and are expected to.
I also use Claude Code a lot more for thinking or writing these days so there's less separation between topics than there are in separate repos.
At this point, I'd argue any "automated memory" is a failure over time. I don't know if I'll be convinced otherwise.
If one were to go full automation, then I'd want generations of what to prune ("I haven't used this memory in a while... "). Actually, that might be useful in your tool as well. "Remove-gates"?
Dislike READMEs written by or with the LLM.
They're filled with tropes and absolutes, as if written by a PdM to pitch his budget rather than for engineers needing to make it work.
At some point, reading these becomes like hearing a rhythm in daily life (riding subway, grinding espresso, typing in a cafe) and a rhythm-matched earwig inserts itself in your brain. Now you aren't hearing whatever made that rhythmic sound, but the unrelated earwig. Trying to think about your README or AskHN, instead thinking about LLM slop and confabulation and trust.
(Just like new account ChatEngineer's comment opening with “The [hard thing] is real.” which leaves the otherwise natural rest of comment more sus.)
Unsolicited opinion: Shouldn't be called total-recall, but curated-recall or selective-recall (prefer this one) or partial-recall or anything that doesn't confuse users or LLM as to (trope incoming, because where did they learn tropes anyway?): what it is (and isn't) …
I have no idea if OPs project is good, but the idea is.
How to use Total Recall. It explains all of the concepts, but lacks a "quick start" section for someone who is ready to get up and running.
How to Learn Total Recall in Two Weeks:
also, how consistent is the model at applying the five-point write gate? "does this change future behavior" feels clear in theory, but in practice i'd expect the model to be overly conservative (dropping things that matter) or overly permissive (saving trivia that felt important mid-session) depending on context. have you noticed any patterns in what slips through vs what gets filtered incorrectly?
For teams that do want shared context, I’ll document a “team mode” gitignore pattern that commits only selected registers (e.g. decisions/projects) while keeping daily logs + preferences/people local.
I've been using it since yesterday, it is doing it work as intended without getting in the way. So far, works great.
Random Tuesday thoughts. Neat work!
One thing I've noticed: the "distillation" decision (what deserves L2 vs stays in L1) benefits from a second opinion. I have my fiscal check summaries against a simple rubric: "Does this change future behavior?" If yes, promote. If no, compact and discard.
Curious if you've experimented with read-time context injection vs just the write gate? I've found pulling relevant L2 into prompt context at session start helps continuity, but there's a tradeoff on token burn.
Great work — this feels like table stakes infrastructure for 2026 agents.