Medhara
The memory and governance core — persistent recall, capability enforcement, and deterministic lineage.
Why agents need memory and governance
- ●Agents are stateless by default — every session starts from scratch, discarding all prior observations, decisions, and patterns.
- ●Retrieval systems return nearest neighbors, not contextually relevant memory — leading to redundant or irrelevant recall.
- ●There is no standard access control for agent memory — any agent can read any data in the shared context.
- ●Debugging agent failures requires re-running the full chain with no guarantee of identical outputs.
What Medhara provides
User login frequency decreased by 34% after policy update
Fraud detection accuracy improves 12% when combined with behavioral signals
Route high-risk transactions through dual-agent verification pipeline
Correlation detected between API latency spikes and false-positive rate increases
Agent retry rate spikes during peak trading hours (14:00–16:00 UTC)
PCI compliance requires masking of all card fields in agent memory
Provenance Chain
Source Events
Linked Workflows
Structured memory with provenance
Typed memory objects with versioning, TTL, scope, and provenance chains.
User login frequency decreased by 34% after policy update
Fraud detection accuracy improves 12% when combined with behavioral signals
Route high-risk transactions through dual-agent verification pipeline
Correlation detected between API latency spikes and false-positive rate increases
Agent retry rate spikes during peak trading hours (14:00–16:00 UTC)
PCI compliance requires masking of all card fields in agent memory
Provenance Chain
Source Events
Linked Workflows
Crystallization pipeline
Raw observations are transformed into structured facts and distilled intuitions.
Memory Crystallization Pipeline
Transforming raw signals into durable knowledge
Raw Observations
User login frequency decreased 34% post-policy update
API response time spiked 2.3× during batch processing window
Retry rate in fraud-detector correlates with peak trading hours
Structured Facts
Fraud detection accuracy improves 12% with behavioral signals
PCI compliance mandates masking of all card fields in memory
Distilled Intuitions
Correlation between API latency spikes and false-positive rate
Agent autonomy correlates inversely with error rate in regulated domains
Capability-based governance
Access to tools and actions is governed through deny-by-default capability policies.
Agent Permissions
fraud-detector v2.4.1 · Capability-based access control
Allowed Tools
Denied Tools
Policy Evaluation Trace
Full decision lineage
Every decision links to its trigger, data sources, policies, and outcomes.
Lineage DAG
Selected Node
Node Metadata
Connected Nodes
Policy-bound agents
Capability enforcement ensures autonomous systems operate within explicit boundaries.
Governance Flow
Policy-bound agent execution ensures autonomous systems operate within boundaries.
Autonomous systems require explicit boundaries.
Where it sits in the stack
PlantoOS Architecture
The stack relationship between apps, agents, runtime, and systems.
Applications
Products and workflows
Agents
LLM-powered autonomous units
PlantoOS Runtime
Execution · orchestration · control
Capability Layer
Policy enforcement and tool access
Medhara Core
Memory · governance · lineage
Enterprise + Public Systems
Databases, APIs, infrastructure
A new compute layer for systems operated by agents.
Key workflows
How data flows through the system in typical usage patterns.
Workflow 1
Input
Agent observation stream
Core Process
Medhara writes typed, versioned memory objects with provenance
Output
Persistent structured memory available across sessions
Workflow 2
Input
Agent recall request
Core Process
Medhara ranks by relevance + recency + access patterns
Output
Contextually accurate retrieval, not just nearest-neighbor
Workflow 3
Input
Policy evaluation request
Core Process
Six-layer stack evaluated, deny-by-default, with audit event
Output
Governed action with full provenance trail
Measured outcomes
↓ 20–35%
Redundant retrieval reduced
↑ 18–28%
Contextual recall accuracy
↓ 30–45%
Debugging time with replay
Indicative ranges from internal benchmarks and early deployments; results vary by workload, model, and infrastructure.
How it integrates
SDK-first integration — governed from the first line of code.
Install SDK
pip install medhara or npm install @medhara/sdk
Initialize client
Connect to your Medhara instance with your project key
Write & recall
Use typed memory operations in your agent loop
from medhara import MedharaClient
client = MedharaClient(project="my-project")
# Write a typed memory object
client.memory.write(
content="Q3 revenue forecast revised to $4.2M",
memory_type="observation",
scope="finance-team",
ttl_days=90
)
# Recall with context
results = client.memory.recall(
query="latest revenue projections",
limit=5,
rank_by=["relevance", "recency"]
)
# Check lineage
lineage = client.lineage.trace(action_id="act_abc123")