Evolution
Corrections that recur across sessions become memory proposals or skill suggestions.
Evolution learns from how you work: corrections and preferences that recur across sessions become memory proposals or skill and instruction suggestions. It runs inside the Dream job (no timer of its own) and is off by default.
evolution setting:
| mode | does |
|---|---|
off (default) | nothing: nothing is recorded, no model is called |
observe | notes which corrections and preferences recur. Keeps no text (only a fingerprint per idea) and calls no model |
suggest | also keeps the text of what recurs, and asks the model to turn it into memory proposals (always to review, whatever the category policy says) and skill or instruction suggestions |
auto-safe | like suggest, but memory proposals follow their category policy: auto ones are saved, review ones wait, never ones are dropped |
auto-safe never applies anything but a memory: skills and instructions always wait for you, and Evolution never touches code or configuration.
What it sees
It reads the corrections and preferences in the session digests of a Dream run (for example "Actually use pnpm instead of npm", "Always answer in Spanish"). Each becomes a fingerprint, the sorted content words of the sentence without the marker words, so "Don't use npm, use pnpm" and "Use pnpm instead of npm" count as one idea. An idea recurs when it appears in at least two different sessions of the project. Evidence is kept 30 days, at most 500 rows per project, and goes with its session when the session is deleted. Text that may contain a secret or credential is never kept.
What it proposes
In suggest and auto-safe, one extra model call (inside the Dream run's token budget) gets the recurring ideas and what is already remembered, and answers with at most 5 candidates:
| kind | becomes |
|---|---|
memory | a project memory of at most 500 characters, with a category (see the policy rules above) |
skill | a suggestion: name, description and a Markdown body |
instruction | a suggestion: text for the project's instructions file |
workflow (opt-in) | a v2 JSON plan with read-only agent tasks and an independent completion audit |
Every candidate is checked and refused (with the reason in the run's report) when it is malformed, cites evidence it was not shown, may contain a secret, repeats a memory, has a name that is not lowercase-hyphen, or a skill body that carries frontmatter. A suggestion that already exists (pending or accepted) is not filed again, and at most 100 wait per project. The ledger retains at most 200 candidates per project and never evicts an active promotion receipt.
Suggestions
| command | does |
|---|---|
evolution.list (directory, optional status) | { mode, suggestions, observed }: the mode, the last 50 suggestions (newest first) and what Evolution has seen there (corrections, preferences, recurring, counted in distinct ideas) |
evolution.resolve (suggestionId, decision) | accept or reject. Accepting a skill writes .mendcode/skills/<name>/SKILL.md in its project and replies with its path; an instruction is only marked accepted (its text is in the suggestion) |
evolution.promote (suggestionId, expectedHash, policyRevision) | Publishes a reviewed skill or workflow; returns { suggestion, path } with a durable receipt |
evolution.rollback (suggestionId, expectedHash, expectedRevision) | Retires the exact promoted artifact, increments its receipt revision and marks the candidate rolled_back |
In /memory, 7 opens Evolution. ↑↓ selects a candidate, pgup/pgdn scrolls its full content and receipt, a promotes (or accepts instruction text), x rejects, r rolls back, ←→ changes mode and o toggles evolutionWorkflows (off by default). Keys stay in the page header. Below 70 columns the panel grid becomes one column.
Workflow candidates require evolutionWorkflows: true and suggest or auto-safe. They use v2's plan schema: every task is agent with mode: read, permission: ask, and an inline contract with audit: independent and no host checks. Promotion writes a new .mendcode/workflows/evo-<name-prefix>-<id>.json for the existing workflow loader; it never starts a run. Global/user definitions, code, configuration and instruction files are not promotion destinations.
Each candidate has a SHA-256 review hash and a policy revision. Changing mode or workflow consent invalidates older candidates, even after switching off and on again. evolution.list includes the current workflows switch and policyRevision when enabled or candidates exist. evolution.resolve also accepts an optional expectedHash; accepting skills/workflows follows the same promotion guards. Neither promotion nor rollback is available to read-only remote tokens.
A write-ahead receipt records destination, exact artifact hash, revision and activation time before publication. A failed/interrupted publication stays blocked with its receipt. Rollback remains available after Evolution is off, refuses changed files or symlinks, and moves intact artifacts into <data dir>/evolution/retired/<id>-<hash-prefix>/ instead of deleting them. It recovers interruptions before publication or after retirement; manually changed or partially written artifacts require manual review. Already running workflows keep their frozen plans.
Writing a skill never overwrites a file, and is refused when the project folder is gone or when .mendcode, skills or the skill's folder is a symlink or not a folder. A refused write leaves the suggestion pending.
The run's report has an evolution part: mode, what was observed, what was suggested (memories, skills, instructions and opt-in workflows), how many memories were saved directly, and a note when no model call was made (nothing recurs yet, observe mode, budget used up).
Ported from v1, and not
Ported: the four modes and their semantics (observe keeps no text and calls no model; suggest proposes; auto-safe applies only what a deterministic rule allows), recurring corrections as the evidence, conservative rejection of sensitive text, bounded evidence and candidate capacity, validation of every candidate, accepting a skill as an explicit action that never overwrites.
Ported additionally: opt-in workflow candidates, explicit promotion, persistent review/policy hashes, write-ahead revision receipts and guarded rollback of skills/workflows.
Left out: v1's other evidence sources (tool results, test results), per-project consent files and the remoteProcessing switch (the model call uses the Dream role's model), and a separate daily Evolution schedule (it runs with Dream).