Decision-Grade Artifacts, not ranked lists
Pensiv is the only production system that implements cognitive memory science end-to-end. Every capability is grounded in decades of research. Every output is auditable back to source.
What a Decision-Grade Artifact looks like
Senior decision-makers don't read ranked lists. Neither should your agents. Pensiv returns a Primary Scenario (most likely) and a Stress Scenario (worst credible), with role mapping, decision points, evidence gaps, and recommended action.
Technical review delayed by customer procurement cycle
- Proceed with architecture proposal now, knowing customer won't review until Q3
- Wait for procurement clarity before committing engineering resources
- Offer interim technical proof-of-concept to maintain momentum
- Customer's actual budget approval timeline (last mentioned in March 14 call with VP Eng)
- Whether technical POC would bypass procurement gate or still require approval
- IT Security's stance on cloud-hosted POC vs on-prem requirement
Budget freeze due to broader enterprise IT consolidation
- Pivot to cost-neutral architecture leveraging existing infrastructure
- Park opportunity until budget cycle resets in FY27
- Escalate to executive sponsor to protect budget line
- Whether IT consolidation is enterprise-wide or division-specific
- Executive sponsor's current influence level after Q2 reorg
- Timeline for consolidation completion (affects when normal budgeting resumes)
Why this shape wins
Separates likelihood from consequence
A single score hides the axis senior decision-makers unpack manually under time pressure. Pensiv emits both. The Primary Scenario is most likely; the Stress Scenario is worst credible. Different questions.
Imposes a plan structure the consumer cannot skip
decision_points and evidence_gaps are required fields. The human or the agent can't read the output and still not know what to do next.
Auditable by construction
The confidence decomposition and derived_from make every artifact paste-ready for a postmortem, regulatory filing, or agent action log. No "trust me" synthesis.
For engineers: Full JSON schema
›{
"artifact_type": "decision_grade",
"scenarios": [
{
"type": "primary",
"confidence": 0.73,
"title": "Technical review delayed by customer procurement cycle",
"role_mapping": {
"customer": [
{"role": "VP Engineering", "authority": "decision"},
{"role": "Director IT Security", "authority": "technical_veto"},
{"role": "Procurement", "authority": "budget_gatekeeper"}
]
},
"decision_points": [
"Proceed with architecture proposal now",
"Wait for procurement clarity",
"Offer interim technical POC"
],
"evidence_gaps": [
"Customer's actual budget approval timeline",
"POC bypass potential",
"IT Security stance on cloud vs on-prem"
],
"recommended_action": {
"action": "Offer interim POC with 2-week delivery",
"historical_outcome": "4/6 accounts accelerated by avg. 6 weeks",
"risk": "Sunk cost if deal stalls"
},
"derived_from": {
"memory_ids": [412, 387, 322, 301, 289, 267, 244],
"span_months": 18,
"sources": ["call_transcript", "architecture_notes", "account_history"]
}
},
{
"type": "stress",
"confidence": 0.41,
"title": "Budget freeze due to enterprise IT consolidation",
...
}
],
"query": "What's blocking the architecture review?",
"retrieved_at": "2026-05-06T14:32:18Z",
"vault_id": "capture_team_vault_acme_corp"
}
Plug into any agent with one config line
Install the Pensiv MCP server
The Model Context Protocol means any agent — Claude Code, Cursor, Codex, custom builds — can plug into Pensiv without vendor lock-in. The memory layer becomes infrastructure, not a feature of any single AI product.
npm install -g @pensiv/mcp-server
# Add to your MCP config
{
"mcpServers": {
"pensiv": {
"command": "pensiv-mcp",
"args": ["--vault", "my_vault"],
"env": {
"PENSIV_API_KEY": "your_key_here"
}
}
}
}
Your agent now has access to:
-
•
pensiv.remember()— write a memory with surprise-gated triage -
•
pensiv.recall()— BM25 + vector + analogical retrieval -
•
pensiv.decide()— Decision-Grade Artifact synthesis -
•
pensiv.audit()— full provenance chain for any memory
Three tiers of connectors
Connectors are the input interface. The MCP server is the output interface. Pensiv owns the cognitive layer between them.
Individual user acquisition
Every knowledge worker generates these daily. Zero-config onboarding. A calendar invite, the email thread before the meeting, and the Slack follow-up all become connected memory without the user doing anything.
Team-level expansion
Structured operational context — ticket state changes, sprint patterns, architectural decisions buried in PR reviews. The surprise gate filters noise across all streams.
Enterprise contract layer
Ground truth for customer relationships, deal history, support escalation patterns, procurement cycles. This is where multi-user vaults show their value — AE call notes, SA proposals, and support tickets converge into org-level memory.
Seven capabilities. Zero market equivalents.
Decision-Grade Artifact
Not a ranked list. Primary + Stress scenarios with role mapping, decision points, evidence gaps, and recommended actions.
Surprise-gated writes
Inspired by Google's Titans. Only surprising information gets committed. Signal compounds; noise stays out.
Analogical retrieval
Gentner's Structure-Mapping Engine. Finds structural matches with zero keyword overlap. Legal precedent, competitive moves, clinical patterns.
FSRS-6 decay
Memory science, not arbitrary TTL. Memory you return to compounds. Memory you ignore ages gracefully.
Full auditability
Provenance + confidence decomposition on every output. Built for ITAR, HIPAA, SEC 17a-4, ICD 203.
MCP-first
Plugs into Claude Code, Cursor, Codex, custom agents with one config line. Portable memory, no lock-in.
One database
PostgreSQL, not a multi-store nightmare. If your IT team can back up a database, they can back up Pensiv.