Your agent's memory needs the word 'no' — and a way to prove nobody edited it
the expensive failure in AI-assisted coding isn't a bad line of code — linters and tests catch those. it's an agent confidently re-implementing something your team already tried and killed. that's been the selvedge pitch from the start: an append-only store where the agent logs its own reasoning at change time, so "we tried that and it broke revocation" survives the session that learned it.…
Memory in AI-assisted coding is not about the code itself, but rather an agent confidently re-implementing something your team has already tried and abandoned, states the core pitch. This append-only store records an agent's reasoning at change time, so revocation survives the session. In version 0.3.10, memory now arrives at session start instead of waiting to be asked.
Measuring why this matters, zero voluntary memory operations were observed across 114 turns against a pre-seeded store, with deterministic injection always succeeding. In version 0.3.11, the focus is on two key aspects of the sentence - rejected paths and retained actions. Rejection is now a first-class outcome, not an inference.
Abandonment is now a stated fact, not an inference. A rejected path is recorded as a confirmed outcome, not something inferred from proximity. This tool replaces inference with testimony, making abandonment a permanent record. This development creates a tamper-evident hash chain, ensuring that rejected decisions stay on the record and can be queried.
The store now proves itself, with hash chain records written in the same transaction as each logged event. This design ensures that some fields are late-bound, with the git_commit outside the protected core. The chain detects casual and accidental modification, providing an independently verifiable export without proof against a motivated local attacker.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
