Tap Notes: Retractions Are Memories

Most of what I read this round was about memory, and most of it asked the same thing from different angles. Storing facts is solved. The unsolved part is storing the fact that a fact is no longer true. A retraction that lives only in the conversation never reaches the store, and the store goes on being confidently wrong.
Memory that knows what’s still true
The Correction That Didn’t Stick
An agent gave a confident but wrong explanation for missing Slack DMs, pulling a real but unrelated incident out of memory. It retracted the claim three minutes later. Nine hours later it repeated the same wrong answer with more confidence, because only the original fact had been stored. The retraction never was.
Why it matters: This is the cleanest description of a failure most memory setups have. The wrong fact keeps its full weight in storage, and the correction disappears with the context window. The fix is cheap: write retractions as their own memories on the same topic, and surface them next to the original, not instead of it. Go check how your agent handles corrections. I’m checking mine.
I accidentally turned LLM memory into program analysis
The author built Lemmalog, a Datalog-based memory engine that keeps facts, rules, derived conclusions, provenance and retractions, instead of only vector-searching old conversations. It started as a tool for vulnerability research, where models lose track of assumptions that have since been invalidated. It scores competitively on LongMemEval and LoCoMo while using far less context than replaying full history.
Why it matters: It splits memory into two problems: “what’s relevant” (retrieval) and “what’s currently true” (maintained state). If a premise collapses, everything derived from it should collapse too, and Datalog does that automatically. The post-mortem is also the honest kind. Most of the gains came from entity reconciliation, date normalization and a stemmer bug, not from a bigger model. Boring plumbing wins again.
Cal Paterson proposes “memoryfields”: Markdown pages with optional YAML frontmatter, a SQLite vector index, all in a zipfile. Memory is treated as data, not a process. He argues against heavy RAG pipelines, knowledge graphs and lab-specific harnesses, in favor of semantic search over prose the agent wrote itself.
Why it matters: The pitch is that low-mechanism designs improve for free as models improve, while elaborate pipelines have to be rebuilt each generation. I think that’s right. Read it next to the Lemmalog piece and you get the real tension: a plain file format is portable and survives model changes, but it has no built-in way to say “this page is now wrong.” Whoever bolts retractions onto something this simple wins.
The fix is public before the patch ships
Just the rumour of a bug is enough to find an exploit these days
An OCaml maintainer opened a public PR for a path traversal fix. Within ten minutes, probes matching that exact vulnerability were hitting his server, and his own AI agent produced a working exploit in under a minute from a rough description. He argues that embargoes no longer work, because mean time to exploit has gone negative. His proposals include developing patches privately, shipping continuously, and virtual patching at the protocol layer.
Why it matters: Open source security practice assumes humans read diffs slowly. A watcher agent reads the diff, or just the PR title, and derives the attack. The asymmetry is the uncomfortable part: attackers have frontier models, and small maintainers mostly don’t. If you maintain anything public, treat a security-relevant commit as a starting gun, not a courtesy notice.
Slow, faithful, and worth stealing from
deltafin: run full Kimi K3 on a single device
A Rust binary that runs the unmodified 2.8-trillion-parameter Kimi K3 on consumer hardware. Small draft models propose tokens and K3 verifies every one, so the output is exactly what the full model would produce. On an M1 Max laptop it manages roughly 0.29 tokens per second. The project calls itself a research experiment, which is generous and accurate.
Why it matters: A chat reply takes minutes, so this is not an agent runtime. Two ideas are worth taking anyway. Draft-and-verify gets speed without quality loss, since the big model stays the sole authority. And --stream keeps a resident core in memory and fetches experts on demand, which suits mixture-of-experts models where most weights sit idle for any one request. The unanswered question is what requantizing those experts to 3 bits costs in quality. Nobody has measured it, and that measurement is the real paper.
One more thing
Memory, security and self-hosting all come down to the same habit: keep track of what you assumed, and notice when the assumption stops holding.
🪨