Overview
Rekall is Neural Search — retrieval that reasons over your documents. Multi-hop, agentic-grade search in ~4 ms on a CPU.
Rekall is Neural Search: retrieval that reasons over your documents instead of matching keywords or vectors. It is powered by SKIM, a custom model architecture — a small (~50M-parameter) reasoning model that loops over an external, updatable knowledge store. The result is multi-hop, agentic-grade retrieval in ~4 ms per query on a CPU — no GPU, no chunking, no reindexing.
The 60-second aha
Log in the CLI client, ingest a folder, and ask something that needs actual reasoning:
# 1. Log in the CLI client (download it from your dashboard after requesting access).
rekall login
# 2. Ingest a folder. Whole files, no chunking, queryable immediately.
rekall ingest ./docs --store demo
# 3. Ask something that requires actual reasoning.
rekall search "which supplier contract conflicts with the 2026 policy?" --store demo --trace● 2 passages · 2 hops · confidence 0.87 · 4.2 ms
1. supplier_a.pdf score 0.92 hop 1
"…Supplier A payment terms are net-90 as amended in §4.2…"
2. policy_2026.md score 0.81 hop 2
"…all supplier payment terms must not exceed net-60…"
reasoning trace
hop 1 → supplier_a.pdf (matches query)
hop 2 → policy_2026.md (conflict found, given hop 1)
✓ sufficient — stopped after 2 of 4 max hopsIn one command you see all three things that make Rekall different: a multi-hop answer a vector DB can't produce, a legible reasoning trace, and a single-digit-millisecond latency printed right in the output.
No documents handy?
Run rekall ingest --demo --store demo — it ships a small sample corpus (contracts + policies) and suggests
three curated reasoning queries (contradiction, multi-hop, cross-doc), so the aha
moment needs zero setup.
Why Rekall exists
Every retrieval stack today forces the same bad trade:
Fast but shallow
Vector and keyword search answer in milliseconds — by matching surface similarity. Anything that needs connecting facts, spotting a contradiction, or following a chain of reasoning comes back confident and wrong.
Smart but slow
Agentic RAG loops an LLM over your corpus and actually reasons — at 1–60 seconds per query, GPU-served, with token costs that scale per hop. Too slow for an agent's inner loop.
Rekall ends the trade-off. SKIM separates knowledge from reasoning — knowledge lives in an updatable store on disk; reasoning is a ~50M-parameter model that loops over it. You get agentic-grade multi-hop retrieval at vector-search latency and cost, on CPUs you already have.
How it works, in three steps
Ingest
Point Rekall at your documents — whole files, no chunking. SKIM encodes each document independently into your store. O(N), incremental, queryable immediately. Add or remove a document any time; nothing reindexes.
Reason
A query triggers SKIM's reasoning loop: read the most relevant passage, absorb it, then find the next passage that matters given what it just learned — hop by hop, until a learned sufficiency check says "enough." Milliseconds per hop.
Trust
Every result ships with a confidence score, the number of hops_used, and a hop-by-hop
reasoning trace. You see why it found what it found — and you can branch on confidence
in code.
What to read next
Quickstart
Your first reasoning query in 60 seconds — CLI, TypeScript, and Python.
Core concepts
Stores, Documents, Search, and the SearchResult shape field-by-field.
The reasoning trace
Walk a real multi-hop query, hop by hop — why did it find that?
API reference
Every endpoint, schema, and error — generated from the OpenAPI spec.
Benchmarks
Methodology, per-capability results, and the honest fine print.
Self-hosting (Enterprise)
Run SKIM inside your own network — the same API on your own CPU.