What it is: AgentPrizm (agentprizm.com) is a hosted, "governed" memory-and-skills layer for AI agents that went public on 2026-07-09. It ships two products on one API key: AgentMemory, a persistent memory service over a REST API and a remote MCP server, and AgentSkills, a marketplace of reusable, versioned agent workflows. The through-line, per the launch release, is that an agent should be able to prove what it remembered, why, and when that memory may no longer be true.
Who it's for: Developers and teams wiring memory into agents that touch real work — the named cases are coding agents (preserve architectural decisions across sessions), support agents (recall a customer's history with source citations), sales, and compliance agents (prove a record was deleted on request). If you're already comparing Mem0 vs Zep vs Letta, this is a fourth name for the shortlist — with governance as its wedge.
Should you care: If you want cross-session memory without standing up your own vector store and write-consolidation logic, yes — especially if audit trails and GDPR-style deletion matter to your buyers. If you're a solo builder shipping a hobby agent, the free tier is generous enough to try and the MCP setup is a paste-one-block affair.
What AgentMemory actually does#
AgentMemory is the part that stores facts your agent recalls. Through the REST API (at /api/v1/agent/*) and a remote MCP memory server, the release lists: confidence-weighted facts, fact-validity windows, contradiction handling, container isolation, hybrid semantic + keyword recall, audit receipts, and GDPR-aligned right-to-forget with verifiable deletion.
Most of that vocabulary maps onto problems the memory field already named. Confidence and validity windows attack staleness — the failure where retrieval keeps returning a fact that was true last month. Contradiction handling is the write-time consolidation step (add / update / delete) that separates real agent memory from RAG: memory has to decide what to do when new information conflicts with what's stored, not just append. Audit receipts and verifiable deletion are the governance framing — proof that a recall happened, or that a record is gone. AgentPrizm — founded and led by Gene Avakyan — positions memory as a trust layer for agents doing business work, which is a reasonable read of where the category is heading.
What AgentSkills adds#
AgentSkills is a governed marketplace and procedure layer for reusable agent workflows. You publish a SKILL.md once; agents discover skills by intent; and the platform preserves lineage and blocks secrets or PII before anything is shared. It rides the same API key and the same REST + MCP surface as AgentMemory, and per the homepage it's included free on every plan. If you've read our take on skills vs subagents vs tools, this is the "reusable procedure, versioned and shareable" end of that spectrum, with a governance wrapper.
How to start#
- Sign up on agentprizm.com and get an API key.
- Paste one config block into Claude Code, Cursor, Claude Desktop, or OpenClaw — it works with any MCP-capable agent. Prefer HTTP? Call the REST API directly from any language.
- It's also published as an OpenClaw skill on ClawHub, OpenClaw's public skill registry, if you live in that ecosystem.
Pricing#
- Free: API + MCP access, 10,000 memories and 4,500 recalls per month. AgentSkills included. All recall features (hybrid retrieval, confidence, validity windows, audit trails) are advertised on every plan, free included.
- Paid: Plans exist for larger workloads, higher recall volume, and production teams. AgentPrizm hasn't published exact tier prices — treat the paid plans as "see pricing" until they do.
The honest catch#
Three things to weigh before you route your agents' memory through it.
You're renting your agents' memory. This is a brand-new hosted service with no track record. Every fact your agents accumulate lives in one vendor's store, reachable through one vendor's API. There's no independent portability story yet, so switching costs compound the longer you run — the memory is the moat, and it's theirs, not yours. Ask about export before you commit anything you'd hate to lose.
The free-tier recall cap is easy to blow past. 4,500 recalls a month sounds fine until you do the math on a live agent: that's roughly 150 recalls a day, and a single multi-step task can fire several. A modestly busy support or coding agent will clear that in days, not weeks, pushing you onto paid plans whose prices aren't public. Model your real recall volume before you assume "free" covers production.
"Governed" is a claim you can't benchmark yet. Audit receipts, verifiable deletion, contradiction handling — these are exactly the right features, and also exactly the ones no third party has independently tested on this platform. There's no public LOCOMO-style benchmark for AgentPrizm, and "patent-pending" isn't a performance number. If governance is why you're buying, evaluate it yourself against your own recall and deletion tests before you trust the marketing.
Bottom line: A genuinely interesting entry that leads with governance instead of raw recall, with a low-friction MCP setup and a real free tier. Try it on a non-critical agent, verify the audit and deletion behavior against your own tests, and keep an export plan in your back pocket. The idea that memory is a trust layer is right — just make sure the trust runs both ways.
