Roadmap

This page is kept honest: “shipped” means merged, tested, and covered by the benchmark gate; “next” lists the real reason each item is not done yet. Dated summaries of each pass live in the changelog.

Recently shipped

The 2026-07 performance pass:

• Much faster writes. Bulk import commits in batches with one fsync each — 50,000 documents load in about 2 seconds — and single durable writes now run at the fsync wall, the same physical wall every embedded database hits.
• Deferred cache persistence. Derived cache sidecars are no longer rewritten on every operation; checkpoints may lag the durable log by a bounded tail, and readers heal them lazily. Persistence is an optimization, never a correctness event.
• describe shows the write contract. The live manual now includes each collection’s indexes and validation rules, so an agent knows what a write must satisfy before attempting it.
• Incremental vector sync. vsync skips embeddings that are already fresh, and vsearch/context reuse persisted vectors instead of re-embedding on every call.
• MCP contract hardening. Unknown tool names report “unknown tool”, destructive tools carry real destructive hints, and every tool declares its required parameters in the schema.
• Structured missing-collection errors. explain and diagnose fail loudly on a missing collection with a did-you-mean hint, instead of silently returning nothing.
• Compaction preserves history. Compacting the op-log keeps each surviving record’s original timestamp, so compaction no longer rewrites when things happened.

Next

Each item was deferred for a real reason, listed with it:

1. Warm-open fast path. Opening from a checkpoint is benchmarked but not yet measurably faster than replaying the log — it ships when it actually wins.
2. Paged caches, then a paged store. Every fresh process currently decodes whole cache sidecars before its first read; paging or memory-mapping them is the stepping stone to datasets bigger than RAM — the main gap to production scale.
3. Uniform --json output flag. Today only inspect --json exists; agents parse text for the rest of the CLI.
4. A real neural embedder (ONNX). The bundled embedder is a deterministic baseline; a neural drop-in needs a native-dependency and model-distribution story that doesn’t compromise the dependency-light build.
5. TypeScript binding. A large new surface over the CLI/JSON layer; worth doing once, properly, with real Node tooling.
6. HNSW vector index. Flat exact kNN is correct and fine at current scale; an approximate index must stay a derived, rebuildable artifact like every other binary file.
7. Log-native rollback. Undo expressed in the op-log itself, rather than only in sidecar change files.
8. Splitting the two largest source files. The store and query crates have grown monolithic files; splitting them is a dedicated refactor, deliberately not mixed with behavior changes.

Principles

• The text stays canonical. The NDJSON op-log is the single source of truth. Binary artifacts — caches, vector indexes, future paged checkpoints — are allowed only as derived, deletable, rebuildable files that verify can check at runtime.
• Honest limitations. Every feature ships with a plain statement of its current limit, and this site never claims past it. If the docs and the engine disagree, that’s a bug in the docs.
• Every claim is verified. Behavior is proven by tests, and performance by a benchmark harness with a regression gate — a reintroduced hot-path cost fails loudly. A number you read here was measured, not estimated.