Veil帳

A session with no memory relearns the codebase from zero every time it starts. It rediscovers the constraint you explained last week and retries the fix that already failed. Compaction does not solve this: the summarizer keeps whatever looks important by some fixed heuristic and drops the rest, and the rest turns out to matter more often than the tool admits.

Veil replaces that summarizer with an eviction system that tracks each piece of context on its own schedule, instead of erasing a session in one pass.

npm install -g @engrammic/veil
cd your-project
veil

The package runs local-first on sqlite-vec embeddings. Nothing about your codebase leaves the machine, and it works with the network off.

Most cache eviction runs on a plain age check: old enough, gone. Veil scores memory with FSRS, the algorithm behind spaced-repetition flashcard apps. Each item gets a stability value in days, and retrievability decays from that stability on a power curve rather than a straight line, so an item that keeps getting recalled stays sharp much longer than one recalled once and left alone.

Stability starts low and differs by what the item is: a passing observation opens at roughly thirty minutes, a fact at two hours, a decision at twelve. Pin something as intent and its stability is set high enough that it never decays. Once computed retrievability drops under 0.1, the item is a candidate for eviction, no matter how recently it was written.

The retrievability threshold that decides what is low enough to evict is not fixed. It borrows the AIMD shape from TCP congestion control: three evictions inside sixty seconds counts as thrashing, and the threshold backs off, keeping more in context until things settle. Five idle minutes with no eviction lets it creep back up. If an evicted item gets asked for again, that counts as a miss, and the threshold rises right away instead of waiting out the idle window.

Nobody sets an eviction budget by hand. The threshold finds its own level from how the session is going.

Deciding what a piece of context is worth does not call the LLM. Every item gets a weighted score from five metadata signals: FSRS retrievability for recency, a log-scaled access count for frequency, tag overlap with the current task, whether the item has a pointer into the structural graph, and a cognitive weight that tracks whether the item was present during past successes or failures. The weights are fixed (0.30 for relevance, 0.25 for recency, 0.15 each for frequency, structure, and cognitive weight), and the whole thing runs as arithmetic.

That keeps scoring under ten milliseconds on every turn. An eviction pass never becomes the thing the agent is waiting on, and the score comes out the same twice in a row given the same inputs, which a model call does not.

Veil keeps a record of every attempt against a goal: what the agent did, the target, the outcome (pass, fail, partial, or uncertain), and a normalized fingerprint of the error so the same failure gets recognized as the same failure even when the message text drifts. A convergence monitor watches that record. Five consecutive failures on the same goal, or ten turns with no measurable progress, and it escalates: a warning first, then a callback the harness can act on, then a hard stop if the agent keeps repeating an approach that already failed three times.

Veil packages the same memory engine as an installable CLI; Engrammic runs it as infrastructure. With the failure record in place, an agent stops retrying a fix that already failed and moves to something else instead.