Engraphis Review: Memory That Coding Agents Can Explain

Why coding agents need memory

Every new agent session starts with an empty context window. The agent can read files and search a repository, but it does not know that your team moved from JWT to PASETO, that the deploy branch changed after an incident, or that a workaround exists because of a vendor bug. File search finds code. It does not preserve the reasoning behind the code.

Most “memory” products solve this by stuffing more text into a vector store. That creates a different problem: the agent gets a pile of semantically similar snippets with no boundary, no history, and no way to tell a current fact from a stale one. Engraphis takes the harder route. It treats memory as a governed record with a scope and a timeline.

What Engraphis is

Engraphis is an Apache-2.0 Python service that stores project knowledge in a local SQLite database and exposes it to coding agents through MCP, a REST API, a Python library, and a local dashboard. The core runs offline. Optional hosted services add analytics, scheduled upkeep, and team controls.

It is not a general-purpose vector database, and it is not a hosted chatbot. It is a memory layer for the specific facts an agent needs to do software work: conventions, architecture decisions, constraints, prior fixes, and the reasons they exist.

How it works

Scope before storage

Memories are organized by workspace, repository, and session. Recall can stay inside that boundary instead of collapsing everything into one undifferentiated store. This matters when one developer works across multiple clients or when a team wants a convention to apply to one service, not every project in the account.

Hybrid recall, bounded context

Engraphis combines vector, lexical, graph, and code-aware retrieval, then packs the result to a hard context budget. The goal is not to return the most text. It is to return the smallest useful set of evidence, with the sources attached.

Provenance and abstention

Every returned fact keeps its boundary, history, and recall rationale. If the evidence is insufficient, the engine can abstain instead of inventing an answer. That is a meaningful difference from a black-box RAG pipeline, where a confident-sounding response can hide a weak retrieval.

Temporal history

When a fact changes, the old version is not silently overwritten. The timeline keeps superseded facts visible, so an agent can distinguish “we used to deploy from the release branch” from “we deploy from main after checks pass.” For engineering teams, that history is often as valuable as the current answer.

A dashboard you can inspect

The local dashboard exposes memories, timelines, recall rationale, graph relationships, audit events, workspaces, and system health. You can see what the agent knows and why, instead of trusting a retrieval score.

What we like

Benchmarks: promising, with the right caveats

Engraphis publishes three registered offline fixtures. They separate context size, retrieval quality, and grounded decision checks. The numbers are useful, but they are small fixtures, not a general coding benchmark.

Important: the payload figure is a serialized JSON-shape estimate, not an MCP transport measurement or provider billing total. The project explicitly says these fixtures do not establish general task performance.

Setup and agent support

The local application requires Python 3.10 or newer. The NumPy-only core supports Python 3.9+. A typical setup looks like this:

python -m pip install "engraphis[all]"
engraphis-dashboard

# MCP for a coding agent
python -m pip install "engraphis[mcp]"
engraphis-init
claude mcp add engraphis -- engraphis-mcp

The dashboard opens at 127.0.0.1:8700. The installed service uses a local embedding model of roughly 80 MB on first download, with a deterministic offline fallback. The core Python factory defaults to the offline embedder.

Published setup recipes cover Claude Code, Cline, Codex, Cursor, Windsurf, and Zed. Engraphis is careful to describe those as configuration recipes, not client certification — a level of honesty we appreciate.

Pricing

Trials require email verification and no card. The project is explicit that shared hosted Team memory and hosted MCP are not yet available, so teams should confirm current access before standardizing on it.

Where Engraphis is still early

• Team memory is not fully here. Roles, seats, audit history, and sync exist, but one shared hosted memory across teammates and agents is a longer-term need.
• It is a local service, not zero-config SaaS. Python, a local database, and an agent configuration are part of the deal. That is fine for developers; it is friction for everyone else.
• The benchmarks are small. The registered fixtures are useful evidence of context efficiency and grounded behavior, not proof that Engraphis will improve every coding task.
• Hosted and LLM features are optional and can cost extra. Review with Jev is experimental and BYOK-only in some paths; managed Jev is not yet available.
• It is not a replacement for docs or code search. Engraphis remembers decisions and conventions. It is not a general document store or a full RAG platform.

Who should use Engraphis?

Engraphis is a strong fit for individual developers who use coding agents across sessions, small teams with shared conventions, agencies that need strict project boundaries, and privacy-sensitive organizations that want memory to stay on a local machine. It is a weaker fit for non-technical users, teams that need a fully managed shared memory today, or anyone looking for a plug-and-play cloud knowledge base.

The verdict

Engraphis is the most credible attempt we have seen at making agent memory inspectable rather than magical. The free local core is genuinely useful, the scope model matches how software work is organized, and provenance plus temporal history are real differentiators that most memory tools do not offer.

It is still early. Shared Team memory is not generally available, the published benchmarks are small offline fixtures, and running a local service is a real setup step. But for developers who want their agent to remember the why — and to show its work — Engraphis is an easy recommendation at the free tier.

FAQ