Jikji: one MCP memory layer for agents — 223 ms, 37x fewer tokens

Jikji is a personal memory layer that lets multiple AI agents share one memory over the Model Context Protocol, running a hybrid BM25 + dense + RRF + reranker search on a single SQLite file — with no external vector or graph database — and returning a search in about 223 ms warm.

In short: Jikji unifies agent memory behind one MCP layer, stores text and images in the same vector space for cross-modal recall, and ships as a single portable SQLite file you own.

What exactly makes Jikji different from other memory MCPs?#

The sharpest difference is multimodal recall. Most agent memories handle text only; Jikji puts text and images into the same vector space (Qwen3-VL pooling), so a text query can retrieve a matching image and vice versa — the repo states plainly that general memory services don't offer this. The footprint is light, too: instead of standing up a separate vector or graph database, it keeps facts, revisions, and moderation state in one node:sqlite file and runs cosine search over it. Trust defaults differ as well — training use is off by default, memories are reviewed before indexing, secrets are refused at write time, prompt injection is quarantined, and deletion cascades to a zero-reference purge.

Think of it as several assistants sharing one notebook instead of each carrying their own — whoever writes it down, the next one finds it immediately.

Where do the 223 ms and the token savings come from?#

Jikji runs lexical (BM25) and dense search together on every query, fuses them with RRF, and re-sorts with a reranker. Per the repo's measurements, that end-to-end path (8B embedder -> BM25+dense+RRF -> 8B reranker) is about 223 ms warm. The token savings come from how it recalls: instead of re-injecting the whole chat history, it injects only the top-8 relevant facts, so recall stays around 200 tokens flat. The repo derives roughly 7x / 19x / 37x fewer tokens at 10 / 25 / 50-message sessions. Retrieval quality is reported as top-1 = 1.00 with zero contamination on a Korean long-term-memory eval; wiring in a real embedder (KURE-v1, MIT) lifts top-1 from a lexical 0.25 to 1.00 on the same set.

What should you watch out for? (honest scope)#

Be clear about the assumptions. The 223 ms, the reranker, and the 8B embedder are the optional GPU config; a default checkout runs a GPU-free lexical scaffold that is honestly lower quality (a real embedder swaps in behind the same interface). The eval sets are small smoke tests that guard correctness and regressions rather than measure headroom, and the token figures are derived from measured memory sizes with stated assumptions. Throughput is about 10 searches per second — sized for a closed beta (10-20 users, 50-100 bursty). So the honest claim is not "fastest in the world" but rather: it delivers this latency, quality, and ownership from a single self-hosted file with no per-call external API tax.

Note: the figures above are Jikji's own measured/derived numbers from the public moche-ai/jikji repo (as of 2026-07), and the repo publishes how to reproduce them in BENCHMARKS.md. It is early-stage (a local loopback MVP), so the numbers may change.

Reference links

• Jikji repository (moche-ai/jikji)
• Jikji benchmarks — method & numbers
• Why Jikji — design rationale