Mneme

Adds automatic memory extraction from raw text. Pass any text — a chat transcript, a journal entry, a meeting note — and Claude returns structured memories that mneme persists. Local-first thesis unchanged: BYO Anthropic key, the mneme infra never sees plaintext or the API key.

```ts
import { Mneme } from '@mnemehq/sdk'
import { ClaudeDistiller } from '@mnemehq/distiller-claude'

const mneme = await Mneme.open({
passphrase: '...',
distiller: new ClaudeDistiller({ apiKey: process.env.ANTHROPIC_API_KEY }),
})

await mneme.distill(`
Had a great espresso on Brick Lane. I really prefer single-origin to blends.
Sarah from marketing replies super fast on Slack DMs — should make that the default.
`)
// → mneme.remember() called for each extracted memory automatically
```

What ships

| Package | From → To |
| --- | --- |
| `@mnemehq/sdk` | 0.1.1 → 0.1.2 |
| `@mnemehq/distiller-claude` | — → 0.1.0 (new) |
| All other packages | unchanged |

Architecture highlights

• SDK-level adapter, not a protocol verb. Distillation fans out to N `remember()` calls under the hood. The wire protocol stays at v0.1.
• BYO Anthropic key, no proxying. Mneme infra never holds your key, never sees your prompts, never sees your responses.
• Tool-use, not JSON.parse. Claude returns a `tool_use` block validated by Zod — no malformed-bracket disasters.
• Anthropic prompt caching enabled by default. System prompt + few-shot examples (~2-3k tokens, stable) cached server-side. Subsequent calls within ~5min pay ~10% of input token cost. Distill 20 conversations for the price of 1.5.
• Production-grade error surface. Retries 429 / 5xx with exponential backoff. Hard cost ceiling via `maxCostUsdPerCall`. Per-call `costUsdEstimate` returned. Observability hook for logging.

What it extracts

Six categories aligned to the protocol `MemoryKind` enum: `fact`, `preference`, `event`, `relationship`, `context`, `skill`. The system prompt refuses third-party gossip, hypotheticals, fleeting moods, and questions. Strict per-memory limits (single self-contained sentence, third-person voice, ≤500 chars).

Quality

• 137 unit tests pass / 2 skip / 0 fail (+19 new tests this release: 7 SDK + 12 distiller-claude)
• Lint clean across 96 files
• Typecheck clean across 7 packages
• `prepublishOnly` chain (build + verify-package + publint --strict) ran clean on both publishes
• Zero warnings on either `npm publish` — compare to the v0.1.0 incident that prompted the safety net in v0.1.1

Coming next

• `@mnemehq/distiller-openai` — same interface against the OpenAI SDK. ~2 hours of work; next minor release.
• `@mnemehq/distiller-local` — on-device extraction via llama.cpp / transformers.js. Gated on ADR 0004 §4 cleanup.
• Eval harness (ADR 0013) — curated corpus + queries with expected top matches, precision@1 across prompt revisions. The artefact that turns "it extracts" into "it extracts measurably better than alternatives".
• Streaming distillation for long transcripts (v0.2).

Postmortem references

• Full design rationale: `decisions/0012-distiller-architecture.md`
• Publish safety net (still in effect): `decisions/0011-npm-publishing.md`

🤖 Generated with Claude Code