# AgentPrizm

> AgentPrizm is the memory of record for AI agents — a hosted agentic-memory layer that gives agents persistent, governed memory via a REST API and an MCP server, so they stop being amnesiac and start actually knowing the people and projects they work with. "Memory of record" means the trusted, auditable source of truth your agents can be held accountable to: every memory has a type, a confidence score, a validity window, and an audit trail, and can be forgotten on request.

**Website:** https://agentprizm.com
**Entity:** VUGA Enterprises LLC (d/b/a AgentPrizm)
**Tagline:** The memory layer agents remember you by

---

## What is Agentic Memory?

Agentic memory is persistent, cross-session memory for AI agents — the ability to store, recall, and govern facts, lessons, preferences, and decisions across conversations instead of starting from zero each session. AgentPrizm is a hosted agentic-memory layer (also called an AI agent memory layer or long-term memory for LLM agents). Without agentic memory, every conversation begins with a blank slate; with it, agents accumulate knowledge, follow standing directives, and build context over time.

---

## What AgentPrizm Does

AgentPrizm stores, retrieves, and governs memories for AI agents. Instead of losing context at the end of every conversation, agents call the AgentPrizm API to:

- Bootstrap every session with the user's full context (name, active projects, open tasks, directives, lessons)
- Recall semantically relevant memories using hybrid vector + keyword search with confidence scores
- Store new memories — facts, lessons, directives, preferences, contacts, bookmarks — as structured records
- Forget memories with full audit trails (soft or hard delete, right-to-forget compliance)
- Ingest URLs and files (PDF, TXT, MD, HTML) into searchable memory chunks
- Extract discrete memories from full conversation transcripts
- Answer questions via RAG: hybrid recall + LLM response with citations

AgentPrizm is a memory layer for AI agents that emphasizes governed AI memory: confidence scores, fact-validity windows, contradiction handling, an audit receipt on every recall, and one-call right-to-forget (GDPR-aligned).

---

## Key Concepts

### Containers
Containers are named scopes that group memories by project, context, or agent. A memory can belong to multiple containers. Example: memories tagged `["trading-bot", "infrastructure"]` are retrieved when either container is queried. Containers prevent cross-contamination between unrelated projects.

### 6 Memory Types
Every memory has a type that controls how it is weighted, displayed, and used:

| Type | Purpose | Example |
|------|---------|---------|
| `fact` | Objective information | "The API rate limit is 1,000 req/min" |
| `lesson` | Experiential learning | "MongoDB $lookup is slow on large collections" |
| `directive` | Rules the agent must follow | "Always run tests before committing" |
| `preference` | Personal or style choices | "Prefers concise responses" |
| `contact` | People and organizations | "Sarah Chen — CTO at Acme, met at conference" |
| `bookmark` | References and URLs | "Useful article on vector databases: ..." |

### Hybrid Semantic + Keyword Recall
Recall combines vector similarity (OpenAI embeddings) with MongoDB text search. Results include a `similarity` confidence score. Filters: container, type, severity, tags, date range, validity windows.

### Validity Windows + Supersedes
Memories carry optional `validFrom` / `validUntil` timestamps. Expired memories are excluded from recall. A memory can declare a `supersedes` relation to replace an older record, preserving the audit chain.

### Confidence Scores
Each memory stores a `confidence` field (0.0–1.0). Recall results return a computed similarity score. Agents can store low-confidence memories and update them as certainty increases.

### Audit Trail + Right-to-Forget
Every write and delete is logged. The `/forget` endpoint supports soft delete (memory hidden from recall but preserved in audit) or hard delete (GDPR right-to-erasure). Audit records track which agent performed each action.

### Conversation Extraction
POST a full conversation transcript to `/conversations` and AgentPrizm automatically extracts discrete, structured memories from it — no agent-side parsing required.

---

## API Surface

**Base URL:** `https://agentprizm.com/api/v1/agent`
**Auth:** `Authorization: Bearer ap_YOUR_API_KEY`

| Endpoint | Method | Purpose |
|----------|--------|---------|
| `/bootstrap` | GET | One-shot context: owner info, directives, lessons, recent containers, top memories |
| `/recall` | POST | Semantic + keyword search with filters |
| `/memories` | GET, POST | List + create memories |
| `/memories/:id` | GET, PATCH, DELETE | Single-memory CRUD |
| `/forget` | POST | Soft or hard forget with audit trail |
| `/conversations` | POST | Extract memories from conversation transcript |
| `/context` | POST | Token-budgeted context block for prompt injection |
| `/profile` | POST | Executive summary per container |
| `/stats` | GET | Usage stats + plan limits |
| `/ingest/url` | POST | Fetch, chunk, and embed a URL into memory |
| `/ingest/file` | POST | Upload PDF/TXT/MD/HTML for chunking and embedding |
| `/ask` | POST | RAG: recall + LLM with citations, SSE streaming |
| `/similarity` | POST | Cosine similarity between two memory embeddings |

### MCP Memory Server
AgentPrizm also exposes an MCP (Model Context Protocol) server — an MCP memory server — with two transports:

**Remote HTTP (production path):** `POST https://agentprizm.com/api/mcp`
- Bearer auth, stateless, JSON response mode
- No local install — agents connect with URL + key

**Stdio (local/dev):** run `npx tsx memory-server.ts` from the AgentPrizm repo
- Reads `AGENTPRIZM_API_KEY` and `AGENTPRIZM_API_URL` env vars

**MCP client config (Claude Code, Cursor, Claude Desktop):**
```json
{
  "mcpServers": {
    "agentprizm-memory": {
      "type": "http",
      "url": "https://agentprizm.com/api/mcp",
      "headers": { "Authorization": "Bearer ap_..." }
    }
  }
}
```

**8 MCP tools:** `memory_bootstrap`, `memory_recall`, `memory_create`, `memory_forget`, `memory_ingest`, `memory_ingest_url`, `memory_context`, `memory_profile`

---

## Pricing

| Plan | Price | Memories | Recalls/month (included) | Agents/Containers | Audit log |
|------|-------|----------|--------------------------|-------------------|-----------|
| Hobby | $0/forever | 1,000 | 4,500 recalls/mo (hard cap, returns 429 when exceeded) | 2 | 7-day |
| Builder | $29/mo | 250,000 | 25,000 recalls/mo included, then $0.40/1k | 10 | 90-day |
| Scale | $249/mo + usage | 1,000,000 | 225,000 recalls/mo included, then metered | ~unlimited | 1-year |
| Enterprise | Custom | Custom | Custom | Custom | Custom |

API access is available on all tiers including Hobby. 14-day free trial on paid plans.

---

## How AgentPrizm Compares

Compared with other AI memory tools (such as Mem0, Zep, Letta, Cognee, or Supermemory), AgentPrizm's distinguishing focus is governed AI memory: confidence scores, fact-validity windows, contradiction handling, an audit receipt on every recall, and one-call right-to-forget (GDPR-aligned). AgentPrizm is also framework-agnostic — it works with any agent or LLM via REST or MCP, with no proprietary SDK required.

---

## Key Links

- Home: https://agentprizm.com
- API Reference: https://agentprizm.com/api-reference
- Documentation: https://agentprizm.com/docs
- Pricing: https://agentprizm.com/pricing
- Solutions for Coding Agents: https://agentprizm.com/solutions/coding
- Solutions for Sales Agents: https://agentprizm.com/solutions/sales
- Solutions for Support Agents: https://agentprizm.com/solutions/support
- Solutions for Legal Agents: https://agentprizm.com/solutions/legal
- Security: https://agentprizm.com/security
- Changelog: https://agentprizm.com/changelog
- Blog: https://agentprizm.com/blog
- About: https://agentprizm.com/about
- Contact: https://agentprizm.com/contact
- Sign Up: https://agentprizm.com/signup