smart_recall:
brief your AI before it starts

Every session with an AI coding assistant starts fresh. There's no carry-over from yesterday's migration, no memory of the config quirk you documented last week, no recollection that the  same package version mismatchburned two hours on a different branch. The AI is capable — but it's always starting from zero.

The cost is invisible because it compounds slowly. A two-minute re-orientation here, a five-minute re-derivation there. Multiply by every task, every developer, every session — and you're paying a significant tax on knowledge your team already paid to acquire.

smart_recall is cachly's pre-task lookup. Before your AI reads a single file, it queries the Brain with the current task description and pulls back the most relevant lessons — ranked by confidence and recency.

smart_recall(
  query: "add Redis caching layer to the auth service"
)

// → 3 lessons retrieved (confidence ≥ 0.72):
//
// [0.91] PATTERN: auth-service uses custom connection pooling;
//        standard ioredis defaults cause pool exhaustion under load.
//        FIX: set maxRetriesPerRequest: null, enableOfflineQueue: false.
//        Confirmed 3×.
//
// [0.84] GOTCHA: AUTH_REDIS_URL env var is set per-deployment;
//        local dev uses REDIS_URL. Always check both in config.
//        Last seen: 2026-05-28.
//
// [0.72] DEPENDENCY: Redis caching here interacts with session store.
//        Test session invalidation after adding cache layer.

Those three lessons take milliseconds to retrieve and seconds to apply. Without the Brain, your AI would need to discover the pool exhaustion issue through trial and error — or you'd discover it in production.

The recall is semantic, not text search. Lessons tagged "redis connection pooling" surface when your query is "add caching to auth service" — because the Brain understands that those are related, even without the exact words overlapping. The relevance is scored by weighted term overlap across lesson content, tags, and historical accuracy signals.

This matters because developers don't file tickets in the exact format lessons were saved. Your team writes naturally; the Brain matches intelligently.

Every lesson returned by smart_recall carries a confidence score. High-confidence lessons (0.85+) are ones the team has confirmed multiple times — they go directly into the AI's working context as facts. Medium-confidence lessons (0.60–0.84) are surfaced as things to watch for. Low-confidence (below the threshold) are filtered out to avoid noise.

Confidence moves. A lesson that worked four times in a row sits near 0.91. One that got invalidated by a framework upgrade erodes toward 0.05 as the Brain sees contradicting outcomes. The briefing you get tomorrow reflects what your team learned today.

smart_recall is a Brain MCP tool — which means any AI tool connected to cachly gets it. Claude Code fires it automatically at session start. The OpenClaw Brain Bridge exposes it as middleware. The VS Code extension surfaces the briefing in the sidebar. The same Brain, the same lessons, regardless of which tool the developer is using that day.

The compounding value isn't dramatic in any single session — it's three minutes saved here, a caught mistake there. But across a team of five developers over six months, those minutes add up to days. And the bugs your AI doesn't introduce because it was pre-briefed with known gotchas — those are the most valuable saves of all.

That's what cachly is built around: compounding knowledge that your team already paid to learn, surfaced automatically so you never pay that tuition twice.