Hybrid Search That Actually Works.
Vector search alone fails at scale. ExoVault combines four engines to find what actually matters — even on encrypted data.
Everyone Has Vector Search. ExoVault Has Four Engines.
Mem0, Zep, LangMem — they all do semantic vector search. So does ExoVault. The difference is what happens when vectors aren't enough. Your agent searches for 'PKCE auth flow' and vector search returns generic results about 'login systems.' Or you search for an error code and get vaguely related debugging notes. Vector similarity is a great starting point, but it's not the whole answer. ExoVault layers three more engines on top — so your agents find the right memory, not just a similar one.
"rate limit" → scores by frequency & rarity
"throttling" → finds "rate limit" by meaning
Exact match without decrypting on server
Follows connections to related memories
Memory ranked #1 in multiple engines scores highest. Near-duplicates removed by MMR.
What the Other Three Engines Do
On every query, ExoVault runs four engines in parallel: (1) Semantic vectors — the baseline everyone has. Find meaning, not just keywords. (2) BM25 keyword scoring — inspired by QMD. Catches the exact technical terms that vectors blur away. 'PKCE' matches 'PKCE', not 'authentication generally.' (3) Blind Index — encrypted token matching that works without decrypting anything. The only way to do exact search on E2E encrypted data. (4) Knowledge Graph expansion — follows links between memories. Find one fact, discover the decision that led to it and the task that depends on it. Then Reciprocal Rank Fusion merges all four ranked lists into one result. A memory that scores well across multiple engines ranks higher than one that only matches on vectors.
Reciprocal Rank Fusion
Four engines return four ranked lists. RRF (Reciprocal Rank Fusion) merges them into one. A memory that ranks #1 in BM25 and #3 in semantic scores higher than one that ranks #1 in only one engine. Then MMR (Maximal Marginal Relevance) removes near-duplicates so your agent gets diverse, high-quality results — not five variations of the same fact.
The Only Search That Works on Encrypted Data
Mem0, Zep, LangMem — they all store your data in plaintext on their servers. That makes search easy for them, but terrible for you. ExoVault is E2E encrypted, which means the server never sees your content. So how does search work? Blind index tokens are generated before encryption — allowing exact-match lookups without decryption. Vector embeddings are computed client-side and stored alongside the ciphertext. The server runs similarity search over encrypted indexes. You get the same search quality as plaintext systems, with zero data exposure. No other agent memory solution can do this.
Standing on the Shoulders of QMD
QMD by Tobias Lutke (CEO of Shopify) proved that BM25 + semantic hybrid search destroys grep for knowledge retrieval. His local search engine inspired our approach. But QMD is local-only, single-user, and works on plaintext Obsidian vaults. ExoVault takes the same core insight and extends it for production AI teams: E2E encrypted storage so your IP stays private, multi-agent coordination so Claude, Cursor, and Codex share one searchable vault, multimodal indexing so you can search what was said in a video, and knowledge graph expansion that finds context grep and vectors both miss. Same search philosophy. Built for the real world.
A codex worth keeping.
Free to start. Encrypted always. Connect your first agent in under a minute.