Agent Harness Database
Memory. Context. Actions.
Governed agent harness in one query.

Agents need more than memory — they need context to reason and actions to execute, today stitched from disconnected vector DBs, caches, and audit logs. Areev unifies them in one governed query, with importance scoring, temporal reasoning, biological decay, and cryptographic erasure native.

Start Building →
Native to the Engine

Six primitives. One engine.

These capabilities don't exist in any existing database. Areev was built around them — not on top of them.

Importance Scoring

Vector DBs rank by similarity. Caches don't rank at all. Neither understands that some memories matter more than others.

→ Not in Pinecone, Weaviate, Redis

Temporal Reasoning

When something happened changes what it means. No existing database can answer "what did the user believe about X, as of Tuesday?"

→ Not in Postgres, Elasticsearch, Qdrant

Biological Decay

Memories that aren't reinforced should fade. No database has a native forgetting curve — you'd need background jobs scanning every store.

→ Not in any existing database

Context Assembly Language

One declarative query fuses hexastore, BM25, vector, and temporal recall — engine-optimised, policy-aware, with provenance built in. Application code stops doing retrieval glue.

→ Not in any existing database

Agent Harness

A versioned, governed agent definition — memory, context, tools, LLM config, and policy posture in one signed object. Publish, diff, roll back. Compliance travels with the harness.

→ Not in any existing database

Atomic Transactions

A memory write needs to store the fact, its embedding, its timestamp, and its importance score atomically. With assembled parts, any can fail independently.

→ Requires distributed transactions that vector DBs, caches, and object stores don't support together.

From question to helpful answer.

When someone writes in, Areev runs a CAL query — assembling the agent's harness, its memory, and the catalog of actions — so the model can think it through and reply.

The question

1

Plain words. Unstructured. Human.

Hey, where's my order? It was supposed to arrive yesterday.

Context Assembly

2 · Harness

Memory, context, and actions — one prompt.

Query John's actual intent.
User Who's asking, recent turns.
Memory Preferences, prior orders.
Actions Tools for this harness.

The agent loop

3 · Think + Act

Loops until the goal is met.

  1. Reasons Check the shipment.
  2. Calls lookup_shipment(order_id)
  3. Reads “out for delivery, 6 pm.”
  4. Answers Grounded in the result.

Response stream

4

Grounded answer. Human voice.

Hi John — your order is out for delivery and should arrive by 6 pm today. Sorry for the delay!

One query. Every primitive.

CAL is how agents talk to Areev. Declare what context you need across every grain type — beliefs, goals, events, reasoning — and CAL composes, budgets, and formats it in a single round-trip. No N+1 queries. No client-side stitching.

CAL
ASSEMBLE ticket_context
  FROM
    task:      (RECALL goals ABOUT $session WHERE status IS OPEN LIMIT 1),
    messages:  (RECALL events ABOUT $user RECENT 5 WITH annotate_relative_time),
    workflow:  (RECALL workflows WHERE trigger = "ticket_classification"),
    knowledge: (RECALL beliefs WHERE tags INCLUDE ["product", "support"] LIMIT 10 WITH rerank),
    user:      (RECALL beliefs ABOUT $user WHERE relation IS PREFERENCE WITH conflict_resolution),
    goals:     (RECALL goals ABOUT $user RECENT 3),
    tools:     (RECALL actions LIMIT 10),
    history:   (RECALL actions ABOUT $user RECENT 5 WITH dedup)
  BUDGET 2500 tokens
  WITH provenance
  FORMAT TEMPLATE ticket_brief
Markdown
## Task
Classify the inbound ticket and route it to the right queue within SLA (opened 4m ago)

## Messages
**agent** (4m ago): Did the rate-limit issue start before or after your plan upgrade?
**user** (3m ago): After.
**agent** (3m ago): Got it — can you share the exact error you're seeing?
**user** (2m ago): API returning 429s even after upgrading to Pro yesterday.

## Workflow
classify → (check_plan, check_product) → assign_severity → route → notify

## Knowledge
- Pro plan rate limit: 1000 req/min (0.98)
- Plan upgrades take up to 2h to propagate (0.95)
- Gateway caches tier info 30m after plan change (0.88)

## User
Prefers email · SLA-sensitive account

## Goals
Migrate Starter → Pro · Integrate API to production

## Tools
classify_ticket · route_ticket · escalate_ticket · check_plan_status · get_rate_limit_config

## History
check_plan_status"Pro active since 2026-04-16"
classify_ticket"billing, P2"
route_ticket"billing-eng queue"

The atoms of agent knowledge.

Instead of replaying raw transcripts, Areev stores semantic grains — the compressed, typed residue of every meaningful interaction. 10 grain types cover the full cognitive surface.

BeliefWhat the agent knows as true
0x01
EventTime-stamped occurrences
0x02
StateSession and entity snapshots
0x03
WorkflowMulti-step procedural memory
0x04
ActionCompleted operations and tool calls
0x05
ObservationRaw sensory inputs
0x06
GoalObjectives and pursuit targets
0x07
ReasoningChain-of-thought traces
0x08
ConsensusMulti-agent shared beliefs
0x09
ConsentUser permission records
0x0A
Compliance native

Compliance built in, not bolted on.

Most agent stacks treat compliance as an audit trail you read after something goes wrong. Areev enforces it on the write path — at the engine, before the grain lands.

GDPR HIPAA EU AI Act SOX CCPA LGPD PIPL

Pre-flight Gates

Violations are blocked at the engine, before the grain lands.

Decision Provenance

Every recall carries what it returned, and what it excluded — with reasons.

Stackable Presets

Apply GDPR, HIPAA, and SOX to one tenant. Strictest rule wins.

Bounded Delegation

Human → agent → sub-agent, capped at depth 5. Audited at every access.

Tamper-Evident Export

Merkle-signed envelopes. Verify a grain belongs without trusting the sender.

Cryptographic Erasure

Destroy the key. The grain becomes mathematically unreadable.

Ready to build?

Build agents that think with clarity.

The human mind, as a primitive. Starting now.

Start Building