Docs that ship code.

01Quickstart

This walks you from zero to a working recall in about five minutes. You'll need a free AgentPrizm account (sign up) and Node 18+ or Python 3.10+.

1. Install the SDK

# TypeScript / Node
npm install @agentprizm/sdk

# Python
pip install agentprizm

2. Initialize the client

import { AgentPrizm } from "@agentprizm/sdk";

const memory = new AgentPrizm({
  apiKey: process.env.AGENTPRIZM_API_KEY,
  namespace: "sales-agent-prod",
});

3. Ingest your first memory

await memory.ingest({
  subject: "acme-co",
  text: "Procurement freezes Dec 15 → Jan 5. Renewal owner: Priya.",
  type: "fact",
  source: "call-2026-01-12",
  validUntil: "2026-12-31",
});

4. Recall it

const memories = await memory.recall({
  subject: "acme-co",
  query: "when can we close?",
  k: 5,
});

// memories[0].text → "Procurement freezes Dec 15 → Jan 5..."
// memories[0].confidence → 0.94
// memories[0].validUntil → "2026-12-31"

02Core concepts

Memories

A memory is a single atomic fact, preference, or event tied to a subject (a user, account, project, or other entity). Each memory has a type, source, confidence, and an optional validity window.

Namespaces

Namespaces partition memory across environments and tenants. sales-agent-prod and sales-agent-staging are isolated; recalls in one cannot see memories in the other.

Retrieval channels

Recall blends four channels: semantic (vector similarity), graph (subject relationships), temporal (validity windows + recency decay), and contradiction-aware (newer facts shadow older ones). You can pin or down-weight any channel.

Validity windows

Memories can carry validFrom / validUntil. Recalls outside the window return the memory with reduced confidence and a stale flag. The agent decides whether to use it or re-verify.

03SDKs & tooling

04Self-hosting

The same binary runs in our cloud, your VPC, or fully air-gapped. Three deployment shapes:

• Docker Compose — single-host, sub-30 minute setup. Best for staging and small workloads.
• Kubernetes (Helm) — production HA, auto-scaling, your choice of object store and Postgres.
• Air-gapped — bring-your-own embeddings model, no outbound network, FedRAMP-ready blueprint.

Self-hosting is included on the Enterprise plan. Talk to us for a deployment review.

05Migration guide

The CLI ships with importers for the most common starting points:

agentprizm migrate from-mem0   --src-key=$MEM0_KEY   --namespace=prod
agentprizm migrate from-zep    --src-url=$ZEP_URL    --namespace=prod
agentprizm migrate from-letta  --src-dir=./letta_db  --namespace=prod
agentprizm migrate from-pgvec  --dsn=$PG_DSN         --namespace=prod  --table=memories

Migrations dry-run by default and report what will be ingested before committing. Existing IDs are preserved, validity windows inferred where possible, and a per-row diff is written to ./migrate-report.jsonl.

06Common patterns

Conversation summarization

Run an LLM summary at end of session, ingest as type: "event" with source pointing at the conversation ID. Recall on subject + topic in the next session — confidence-ranked summaries surface first.

Preference learning

Capture preferences as type: "preference" with no validity window. Contradictions resolve newest-wins by default; pass contradictionPolicy: "vote" to require N agreeing memories before flipping.

Tool-call grounding

Before any high-stakes tool call, recall the subject's recent state and inject the top-k memories into the system prompt. The agent stops asking questions it has already asked.