Decisions

Decision memories capture architectural choices, technology migrations, and significant code restructurings. They are the core of what Mindr teaches agents about your codebase.

How decisions are detected

Every commit processed by the post-commit hook is scored against five trigger signals. If the combined confidence score is ≥ 0.15, a decision memory is stored.

Trigger signals

Signal	Weight	Condition
keyword	0.40	Commit message contains a decision keyword
cross-module-diff	0.25	> 100 lines changed across ≥ 2 module directories
new-directory	0.30	A new top-level src/* directory appears
dep-change	0.15	package.json, Cargo.toml, go.mod, or pyproject.toml changes
import-pattern-change	0.25	≥ 5 files add the same new import

Decision keywords (any match in the commit message subject line):

refactor, switch, migrate, chose, decided, architecture, replace,
rewrite, adopt, move to, shift, transition, upgrade, downgrade,
extract, consolidate, standardise, standardize

Confidence score

Confidence is the sum of triggered signal weights, clamped to [0, 1]. A commit that changes a package.json, touches files across two directories, and has the word "migrate" in the message would score 0.15 + 0.25 + 0.40 = 0.80.

Stored decisions expose the full trigger list and confidence score in their metadata.

Manual decisions

Store a decision without a commit:

mindragent remember "Chose PostgreSQL over MongoDB for ACID guarantees" --type decision --module db

Or via the SDK:

await mindr.remember('Chose PostgreSQL over MongoDB', {
  type: 'decision',
  module: 'db',
  metadata: { date: '2026-05-01' },
});

Manual memories have role: 'user' and receive a 20-point quality bonus from the manual-capture component.

Reversing decisions

When an approach is superseded, mark the original decision as reversed:

mindragent decisions reverse <id>

This stores a reversal marker memory. Reversed decisions still appear in decisions list (struck through) but are excluded from context injection. The reversed: true field is set on the Decision object returned by the SDK.

Querying decisions

# CLI
mindragent decisions --module api --from 2026-01-01
mindragent replay --show-reversed   # chronological, with reversals visible

// SDK
const decisions = await mindr.getDecisions({
  module: 'api',
  from: new Date('2026-01-01'),
});

In session context

The top 5 recent decisions (by quality score) appear in mindr:get_context output:

## Recent Decisions
- [2026-05-01] [api] Switch internal APIs from REST to tRPC
- [2026-04-15] [auth] JWT + Redis refresh tokens for access control
- [2026-04-01] [db] Chose PostgreSQL over MongoDB for ACID guarantees

Agents use this to avoid re-litigating past choices and to write code consistent with the architecture.

Decision lifecycle

Commit pushed
    │
    ▼
post-commit hook
    │
    ▼
watcher scores 5 signals
    │
confidence ≥ 0.15?
    │ yes
    ▼
store decision memory
    │
    ▼ (later, if superseded)
mindragent decisions reverse <id>
    │
    ▼
reversal marker stored
decision excluded from context