SEATTLE - s4story -- Lumetra today announced the general availability of Engram, a memory layer for AI agents. After a year of invitation-only beta access, Engram is opening its doors to all developers.

TL;DR

• 91.6% on LongMemEval-S (458/500) out of the box. Methodology and results published openly.
• Every recall is auditable: semantic retrieval plus an automatically maintained knowledge graph, so you can see which memory and which edge produced an answer.
• BYOM by default: frontier, open-source, or self-hosted models. No inference lock-in.
• Plug in three ways: MCP server, REST API, or official TypeScript, Python, and Go SDKs.

What Engram Does

An agent that talked with a user last week recalls their preferences today, and surfaces the exact stored memory and graph edge that produced the answer. Engram ingests conversation, extracts atomic facts and relationships, and stores them where they can be retrieved semantically and explained structurally.

Retrieval fuses three signals: keyword (BM25), semantic vector search, and traversal of the knowledge graph. Recall doesn't fail when a question is rephrased or when the answer depends on an implicit connection between memories.

More on S For Story

• From Luxury Editorial To Lyrical Romance: Editor-in-chief Releases Novel
• TAYP Expands Athlete Exposure Platform Beyond Georgia With New Push Into Virginia and the 757
• KT Medical Staffing Expands Concierge Nursing and Private Duty Nursing Services in Orange County
• The Millennium Alliance Achieves Great Place To Work® Certification™ Amid Continued Growth
• The Millennium Alliance Appoints Former Adweek Executive Eric Hayden Shakun as Chief Financial Officer to Accelerate Next Phase of Growth

For developers, that means memory stops being a black box. The recall path is inspectable end-to-end: every retrieved fact is grounded in a stored memory; every connection is grounded in a graph edge. If a recall is wrong, you can see why.

Memory That Shows Its Work

Memory products that bundle inference and hide the retrieval path require trusting the vendor about what their system did and why. Engram's design choice is the opposite: the system shows its work. This is the difference between a memory product and memory infrastructure.

Bring Your Own Model

Engram is BYOM by default. Developers connect their preferred LLM (frontier, open-source, or self-hosted), and Engram handles extraction, storage, and retrieval. No inference lock-in, no markup on tokens you could have bought directly.

Plug In Anywhere

Engram launches with three integration paths:

• The MCP server works with Claude.ai web, Claude Desktop, Claude Code, Cursor, Windsurf, Codex, ChatGPT, and OpenClaw out of the box.
• The REST API provides standard HTTP endpoints for ingest, query, memory management, and usage stats.
• Official SDKs are available for TypeScript (@lumetra/engram on npm), Python (lumetra-engram on PyPI), and Go.

Pricing

Engram launches with usage-based pricing. No per-seat, no per-project surcharges:

• Free is for evaluation and hobby projects.
• $29 per month covers indie developers and small teams.
• $99 per month covers production workloads.
• Enterprise is custom and includes dedicated support.

Paid tiers meter only on memories stored and retrievals served. There are no per-call inference fees that scale with your success, and no per-token surcharges layered on top of the model you already pay for.

More on S For Story

• T. Jones Group Named Finalist Across Multiple Categories at the 2026 Georgie Awards
• The Simplest Small Business You're Probably Not Thinking About
• San Francisco Writer Wins Webby Award, Internet's Highest Honor, for Website Based on her Novel
• EDC Weekend Comedy Special Featuring Don Barnhart & Friends — Use Promo Code FRIEND for 50% Off
• N Y S E: OTH Off The Hook YS Is Building a Vertically Integrated Marine Empire — And Investors Are Starting to Notice

A Year Behind Closed Doors

Engram spent the past year in invitation-only beta with design partner NeonBay (neonbay.ai), running in production while the team hardened the retrieval pipeline, ingest path, and recall quality. Today's launch opens that same system to every developer.

Quotes

"MCP changed the math on memory. Once a client speaks MCP, adding long-term memory is a config change instead of a rewrite. We built Engram MCP-native from day one because we think that's where the ecosystem is going."
Ben Meyerson, Co-Founder, Lumetra

"Most memory products are black boxes. You hand over your data and trust that the right thing comes back. Engram is built so every recall points to a stored memory and a graph edge. If you don't like an answer, you can see exactly where it came from."
Jacob Davis, Co-Founder, Lumetra

About Lumetra

Lumetra builds memory infrastructure for AI agents. Founded in 2025 by Ben Meyerson and Jacob Davis (previously on AWS IoT at Amazon Web Services), the company is headquartered in Seattle, WA. Engram is its first product.

Start free: https://lumetra.io
Documentation: https://lumetra.io/docs
LongMemEval methodology and results:https://lumetra.io/engram-on-longmemeval