Persistent Agent Memory

Agents That Remember Everything

Store, search, and recall context across sessions in <5ms. Semantic memory is built in — no external vector DB, no extra vendor, no glue code.

HOW IT WORKS

SQL + Vector in One Query

Agent memories are stored as structured rows with embedded vectors. The semantic_match() function enables hybrid queries — join structured data with semantic context in a single round-trip. Memories persist across sessions, are namespaced per agent, and support TTL-based expiry.

-- Recall relevant past resolutions
SELECT t.ticket_id, t.resolution,
  semantic_rank(m.embedding,
    'billing dispute') AS relevance
FROM tickets t
JOIN_VECTOR agent_memories m
  ON semantic_match(m.embedding,
     'billing dispute resolution', 0.75)
WHERE t.customer_id = 'cust_12345'
ORDER BY relevance DESC
LIMIT 5;

WHY IT MATTERS

Why Agents Need This

No More Amnesia

Without persistent memory, every agent session starts from zero. Context is rebuilt, questions are re-asked, and users repeat themselves. Memory fixes this.

Semantic, Not Keyword

Agents need to recall by meaning, not exact strings. semantic_match() finds relevant memories even when the wording is different.

Cross-Agent Sharing

Memories are namespaced but shareable. A support agent can access resolution history from a billing agent — building organizational knowledge.

COMPARISON

Legacy Approach vs ANDI

Legacy Approach

ANDI Approach

External vector DB + custom glue code
Built-in semantic memory
Separate persistence layer per agent
Namespaced, shared across agents
Keyword search only
Semantic search with cosine similarity
No TTL or expiry
Configurable TTL per memory
Two round-trips (SQL + vector)
Single hybrid query via JOIN_VECTOR

Agent-Native Data Infrastructure.
Production-Ready Today.

Persistent memory. Isolated state. Verifiable reasoning. Through standard SQL.