v3.4Now with auditable receipts & fact-validity windows · MCP-native

The memory layer
agents remember you by.

AgentPrizm gives AI agents permanent, governed memory — with confidence scores, audit trails, and fact-validity windows. Drop in one API. Your agents stop starting from zero.

SOC 2 Type II · HIPAA · GDPR · <180ms p95 recall · 100B+ tokens/mo

In production at

LogisticsPrizm
Freight ops platform · agent-assisted dispatch
Powering memory and recall across LogisticsPrizm's dispatch agents in production.
/dashboard/recall/audit/usage
agentprizm.dev / sales-agent-prod
POST /v1/recall · 142ms · 4 results
"What does Rachel care about for the Q3 renewal?"
factm_8f2a · supersedes m_6c910.94
Rachel's procurement needs SOC 2 Type II report before signing. Last renewal blocked 3 weeks on this.
valid: 2026-01-12 →source: zoom_call_2891
preferencem_7e1c0.89
Prefers async updates over standing meetings. Slack > email > calls.
valid: 2025-09-04 →confirmed 4×
eventm_9a550.81
Promoted to VP Eng at Arc on March 14. Now owns infra budget she didn't before.
valid: 2026-03-14 →source: linkedin_diff

/ The problem

Stateless agents are expensive amnesiacs.

Every conversation starts from zero. Users repeat themselves. Facts go stale. Agents contradict last week's answer this week. RAG plus a vector DB doesn't fix it — that's retrieval, not memory.

Without AgentPrizm

Your agent re-introduces itself, every time.

User: "Pull the renewal numbers for Arc."Agent: "Sure — what's Arc? Could you share the account ID?" User (3 days later): "Schedule a call with Rachel."Agent: "Who's Rachel? What email?"User: "...I told you yesterday."Agent: "I don't retain context between sessions."
With AgentPrizm

Your agent knows your business cold.

User: "Pull the renewal numbers for Arc."Agent: "Arc renewal Q3 — $XXXk ARR, blocked on procurement last cycle. Rachel (now VP Eng, owns budget) prefers async. Want a Slack thread or a draft email?"User: "Schedule a call with Rachel."Agent: "She's async-only this quarter — confirmed across 4 prior threads. Want me to draft an agenda doc instead?"

/ Primitives

Memory, not just retrieval.

Six primitives we built because storing-then-searching wasn't enough. Memory has tense, trust, and lineage.

Confidence-weighted facts

Every memory carries a calibrated 0–1 confidence. Below threshold? It asks instead of guessing. Above? It commits with a receipt.

memory.confidence ≥ 0.6

Fact-validity windows

"Rachel was a Senior PM" is true until March 14, 2026. AgentPrizm tracks when facts begin, change, and expire — so old context doesn't poison new decisions.

valid_from / valid_to

Auditable receipts

Every recall ships with: which memories matched, why, what was filtered, and the supersede chain. Compliance, debug, and trust — same primitive.

GET /audit/:trace_id

Container scopes

Per-user, per-org, per-agent, per-session. Isolation by default. Cross-scope queries are explicit and logged. SOC 2 ready out of the box.

ns: user:<id> / org:<id>

Right-to-forget

Hard-delete with cryptographic proof. GDPR Article 17, in one POST. The audit trail records the forgetting itself.

POST /forget

4-channel retrieval

RRF fusion across geometric, lexical, entity-graph, and temporal channels — every channel on every plan, no graph-tier upcharge.

fisher-rao + bm25 + graph + time

/ Drop-in API

Three endpoints. That's the whole product surface.

Ingest a conversation, recall what matters, build context within a token budget. Everything else — extraction, deduplication, contradiction handling, decay — happens for you.

SDKs for TypeScript, Python, Go, Rust
First-class. Not a thin wrapper.
MCP-native — Claude, Gemini, Codex
Auto-inject and auto-capture in your agent runtime.
Self-host on your VPC
Same API, your data plane. For teams that can't egress.
typescriptrecall.ts
import { AgentPrizm } from "@agentprizm/sdk";

const rc = new AgentPrizm({ apiKey: process.env.RC_KEY });

// 1. Ingest a turn — extraction is automatic
await rc.ingest({
  ns: "user:[email protected]",
  role: "user",
  content: "I got promoted to VP Eng. I own the infra budget now.",
});

// 2. Recall — confidence-ranked, validity-aware
const mems = await rc.recall({
  ns: "user:[email protected]",
  query: "what does she care about for renewals?",
  budget: 1200, // tokens
  minConfidence: 0.6,
});

// 3. Drop straight into your prompt
const prompt = `${mems.asContextBlock()}

User: ${userMsg}`;

/ Where AgentPrizm wins

Built for agents in high-stakes work.

Generic memory is fine for chatbots. For agents that close revenue, defend cases, or resolve tickets — you need governance, validity, and proof.

$

Sales agents

Persistent prospect intelligence across reps and renewals. Your SDR-bot knows last quarter's blockers, who got promoted, and which objection you've heard four times.

CaptureCRM events, call notes, email threads
RecallPer-account memory with stage + objection history
GovernanceRep-level namespaces, exportable on offboarding

Support agents

Tier-1 bots that remember the customer, the device serial, last week's workaround, and the policy that just changed. Validity windows = no stale "have you tried restarting" advice.

CaptureTicket history, KB articles, device telemetry
RecallValidity-windowed: stale policies auto-expire
GovernanceSOC 2-ready audit log on every retrieval

Legal & compliance agents

Case agents that cite precedent with provenance. Right-to-forget is a single API call. Every recall is reproducible from its trace ID. Built for regulated workloads.

CaptureFilings, opinions, internal memos with source IDs
RecallReproducible from trace ID — defensible in review
GovernanceRight-to-forget & retention policy via API
180ms
p95 recall at 100B+ tokens / month. Zep clocks ~4s. Mem0, ~7s.
87.7%
LoCoMo benchmark. Top-tier accuracy without graph-tier pricing.
3×
Average context-token reduction vs. naive RAG, on equivalent tasks.
0
Lines of taxonomy you write. We file. You forget about filing.

/ Pricing

Free until you ship. Honest after.

No surprise graph-tier upgrades. Every retrieval channel — geometric, lexical, graph, temporal — is on every plan.

Hobby
$0/forever
For prototypes and side projects. No credit card.
  • 10K memories, 7-day audit log
  • 2 namespaces · 1 environment
  • Community Discord
  • All four retrieval channels
Start free
Builder
$29/mo
For solo devs and indie agents with paying users.
  • 250K memories, 90-day audit log
  • 10 namespaces · 3 environments
  • 500K recalls/mo included
  • Email support · 99.9% uptime
  • All four retrieval channels
Start building
Most teams
Scale
$249/mo + usage
For teams running production agents at scale.
  • Unlimited memories, 1-year audit log
  • Unlimited namespaces & envs
  • SOC 2 Type II + HIPAA BAA
  • p95 SLA · priority support
  • $0.40 / 1k recalls · $0.10 / 1k ingests
Get API key →
Enterprise
Custom
Self-host, custom retention, dedicated infra.
  • Single-tenant or VPC self-host
  • Custom DPAs, residency, retention
  • SSO + SCIM + audit export
  • Solutions engineer + 24/7
  • Volume pricing on usage
Talk to sales

Give your agents a memory
worth keeping.

Six lines of code. Confidence scores, validity windows, and audit trails included. Free until your agents ship.