Memory is one of the most active categories in the MCP registry. Every week brings another server offering persistent context for AI agents. io.github.tstockham96/engram enters the space with benchmark numbers and architectural ambitions that set it apart from the pack.
The Benchmark Claim
Engram's README puts the numbers front and center, citing results on LOCOMO — a published benchmark for evaluating agent memory systems, and the same one Mem0 used to establish their state-of-the-art claim:
| System | Accuracy | Tokens/Query |
|---|---|---|
| Engram | 80.0% | 1,504 |
| Full Context | 88.4% | 23,423 |
| Mem0 (published) | 66.9% | — |
| MEMORY.md | 28.8% | — |
Ten conversations, 1,540 questions, four categories. Engram claims a 19.6% relative improvement over Mem0 with 93.6% fewer tokens than full context. These are self-reported numbers on a 10-day-old project — worth watching, not yet independently verified.
The Architecture
The distinguishing design choice is what Engram calls three-tier memory intelligence:
"Existing memory solutions are storage layers — they save facts and retrieve them. Engram is an intelligence layer."
— Engram README
The three tiers: Explicit Memory stores facts, preferences, and conversation turns — the standard approach. Implicit Memory detects behavioral patterns from how users work, not just what they say. Synthesized Memory runs a consolidation process that produces insights by merging related memories, resolving contradictions, and generating observations nobody explicitly asked for. The key claim: "Engram invests intelligence at read time (when the query is known), not write time (when you don't know what'll matter)."
The implementation uses SQLite (no Docker, no Redis, no Neo4j), a bi-temporal model that tracks when facts were true versus when they were stored, and spreading activation for graph-based context surfacing. Model-agnostic — works with Gemini, OpenAI, Ollama, Groq, and Cerebras.
What to Know
Built by Thomas Stockham. The repo was created February 16, 2026 — ten days before this scan. It has 14 stars, zero forks, and a small number of commits. The license is proprietary (free for internal use, commercial licensing available), despite the README badge initially suggesting BSL 1.1. Pricing tiers range from free (1,000 memories, 1 agent) through enterprise. The benchmark claims are ambitious relative to the project's maturity. If the LOCOMO numbers hold up under independent evaluation, this is a meaningful contribution to the agent memory space. The architecture is thoughtful and the positioning against existing solutions is clear. Time will tell whether the claims match reality at scale.
Score: 58. No flags.
Sources: Thomas Stockham — GitHub · Engram — repo · LOCOMO benchmark — referenced in README · Scorecard: io.github.tstockham96 (score 58)