Context Amnesia
Agents lose context between sessions, forgetting everything they learned.
A Postgres-native memory framework for AI agents. Hierarchical context, multi-agent coordination, zero hard-coded values.
Current AI agent frameworks treat memory as an afterthought. The result? Unreliable, expensive, and impossible to debug.
Agents lose context between sessions, forgetting everything they learned.
Without grounded facts, agents make up information that sounds plausible but is wrong.
40-80% of multi-agent systems fail due to coordination problems and race conditions.
Stuffing entire conversation history into context burns tokens and money.
Most AI frameworks bake in assumptions that don't fit your use case.
When things go wrong, there's no way to trace what the agent was thinking.
A Postgres-native memory framework built from first principles. No magic, no hidden defaults, just clean architecture.
Trajectory ā Scope ā Artifact ā Note. Structured memory that persists across sessions.
="token-keyword">class</span>=="token-keyword">class</span>="token-string">"token-comment"</span>>// ="token-type">Memory</span> hierarchy</span>
="token-type">Trajectory</span> (task container)
āāā ="token-type">Scope</span> (context partition)
ā āāā ="token-type">Turn</span> (ephemeral buffer)
ā āāā ="token-type">Artifact</span> (preserved output)
āāā ="token-type">Note</span> (cross-trajectory knowledge)Every fact backed by stored artifacts. Validation, checkpoints, contradiction detection.
="token-keyword">class</span>=="token-keyword">class</span>="token-string">"token-comment"</span>>// ="token-type">PCP</span> validation</span>
="token-keyword">let</span> checkpoint = pcp::="token-function">create_checkpoint</span>(
trajectory_id,
scope_id,
&artifacts
)?;
="token-keyword">class</span>=="token-keyword">class</span>="token-string">"token-comment"</span>>// ="token-type">Verify</span> grounding</span>
pcp::="token-function">validate_claim</span>(&claim, &checkpoint)?;Locks, messages, delegation, handoffs. Built-in primitives for multi-agent systems.
="token-keyword">class</span>=="token-keyword">class</span>="token-string">"token-comment"</span>>// ="token-type">Agent</span> coordination</span>
="token-keyword">let</span> lock = agents::="token-function">acquire_lock</span>(
resource_id,
agent_id,
="token-type">LockMode</span>::="token-type">Exclusive</span>
)?;
agents::="token-function">delegate_task</span>(
from_agent,
to_agent,
&task
)?;Entity-Component-System pattern for maximum composability. Each crate has a single responsibility.
Bypasses SQL parsing overhead via pgrx. Memory operations happen at native speed.
Provider-agnostic embeddings. Any dimension, any provider. No vendor lock-in.
Compiles to CaliberConfig, not SQL. Type-safe configuration with zero runtime overhead.
Pay only for what you use. No hidden fees, no surprises.
CALIBER is open source under the AGPL-3.0 license. Self-hosting is always free.