Rekalldocs

Changelog

Release notes for Rekall — Neural Search. What shipped, when, across the CLI, REST API, MCP server, and both SDKs.

All notable changes to Rekall are recorded here. Versions are path-additive: the API is version-prefixed (/v1) and additive changes never break clients.

v0.2.0 — July 2026

Closed API positioning. Rekall is now a closed API: the SKIM engine runs server-side and is reached only through the Rekall API. Getting started is request access → API key → SDK or CLI client. The CLI is distributed from the dashboard as an API client (rekall login); the engine no longer runs on the end user's machine.

  • Real 2.6B-token SKIM checkpoint in production. The production Go engine now serves the parity-tested 2.6B-token SKIM checkpoint. (Research-harness figures in the benchmarks remain the PyTorch measurements; production-engine latency is labeled separately and still being optimized.)
  • Versioned model artifacts — REKALL_MODEL_URI points a self-hosted engine at a versioned model folder, from a local path or an s3:// / gs:// bucket.
  • Project-scoped knowledge storage — stores live in per-project subfolders under REKALL_DATA_DIR, with REKALL_PROJECT as the default.
  • Enterprise self-hosting — the licensed binary + model artifact, deployable on Cloud Run, Kubernetes, or VMs. See Self-hosting (Enterprise).

v0.1.0 — July 2026

Initial release. Neural Search, powered by SKIM, across a CLI, a REST API, an MCP server, and TypeScript + Python SDKs. Two engines ship: skim-v1 (the neural reasoning loop) and lexrank (an always-available lexical-hybrid fallback). Every search response reports which one served it.

CLI

A single static binary — an API client for Rekall, distributed from the dashboard.

  • rekall ingest --demo — ingests a bundled 8-document sample corpus and suggests reasoning queries it is built to answer (contradiction, multi-hop, cross-doc).
  • rekall login — browser auth into ~/.rekall/config.
  • rekall ingest <path…> --store <name> — files, folders, globs; --watch, --format auto|text|pdf|md, --strict.
  • rekall search "<query>" --store <name> — --hops, --max-results, --trace, --json, and -i interactive REPL.
  • rekall stores (list / create / stats / delete) and rekall docs (list / delete).
  • rekall serve — the enterprise self-hosted engine (licensed).
  • rekall mcp — stdio MCP server.
  • rekall keys (create / list / revoke).

REST API

OpenAPI 3.1, published at /v1/openapi.json.

  • Stores — create, list, get, delete. Addressable by name or id.
  • Documents — ingest (1–1000 per request, upsert semantics, no reindex), list, get info, delete. Idempotency-Key supported on ingest.
  • Search — the multi-hop reasoning loop. Returns passages, calibrated confidence, adaptive hops_used, an optional hop-by-hop trace, server-measured latency_ms, and the engine that served it.
  • Health — unauthenticated status / version / engine.
  • Structured, prescriptive errors — every error carries type, message, and a fix. Path-versioned /v1; additive changes never break clients.

MCP server

  • rekall mcp exposes stores as MCP tools over stdio — one line of config gives Claude Code, Claude Desktop, and Cursor millisecond inner-loop retrieval.

TypeScript SDK — rekall-search

  • Zero-dependency, pure fetch; runs on Node 18+, Bun, Deno, and edge runtimes.
  • rekall.store(name) get-or-create; fully typed SearchResult / Passage / TraceStep.
  • Streaming ingest (ingestStream) and a Node files helper (ingestFiles).
  • Typed error subclasses carrying the server's fix; automatic jittered-backoff retries on idempotent operations.

Python SDK — rekall-search

  • Sync-first Rekall plus AsyncRekall; dataclass results.
  • search_many for batched offline evaluation; ingest_stream / ingest_files for large corpora.
  • In-package framework adapters: rekall.integrations.langchain.RekallRetriever, …llamaindex.RekallRetriever, and a generic as_tool() for agent frameworks.
  • Typed exceptions mirroring the TypeScript SDK; RekallError.fix carries the remedy.

Self-hosting (Enterprise)

  • rekall serve — the identical API on your own CPU-only box, under an enterprise license. Same SDKs, cloud or on-prem — only REKALL_BASE_URL changes.

Coming next (not in v1)

Explicitly out of scope for v1, tracked for future releases:

  • Managed connectors (Drive / Notion / Slack) — v1 wins on the primitive first.
  • In-store multi-tenancy partitions — the v1 answer is one store per tenant; stores are cheap.
  • Reranking / answer-synthesis endpoints — Rekall returns passages, not answers, by design; pairing with an LLM stays the caller's choice.
  • Rust / Go SDKs — generated from OpenAPI when demand shows.

Benchmarks

Reasoning-grade retrieval at millisecond latency — methodology, macro and per-capability results, and the honest fine print. Numbers are measured; the model is early in training.

On this page

v0.2.0 — July 2026v0.1.0 — July 2026CLIREST APIMCP serverTypeScript SDK — rekall-searchPython SDK — rekall-searchSelf-hosting (Enterprise)Coming next (not in v1)