Engraphis gives AI coding agents a memory that runs on your own machine

A memory layer for coding agents

Engraphis is an open-source project, as github.com reported, that gives AI coding agents persistent memory. It stores knowledge about a project, finds the relevant pieces when asked, and hands back a bounded amount of context along with the sources it came from. The agent, or you, can inspect those sources.

The pitch is simple. An agent that works on a codebase over days or weeks needs to remember decisions, such as why a team changed how login works. Without some kind of memory, that reasoning has to be rediscovered or retyped each time.

Our newsletter flagged it as early stage but worth watching, especially if you run agents through long coding sessions. That is a fair summary of where it stands.

How the memory works

The local engine is built on SQLite, a small database that lives in a single file, and it works offline. Memories are organized by workspace, repository and session, and there is a hand-off from the last session so the next one can pick up where you stopped.

When you ask a question, Engraphis searches in several ways at once. Vector search finds text with similar meaning, lexical search matches the literal words, graph search follows relationships between memories, and code-aware search looks at the code itself. The results are then packed to a hard size limit so the agent is not flooded.

In the project's own example, you store a note that an authentication system moved from JWT to PASETO (two ways of handling login tokens). Later you ask why authentication changed, and the answer comes back with context drawn from that note.

• Install with pip, Python's package tool, then launch a local dashboard at 127.0.0.1:8700.
• The full local application needs Python 3.10 or newer; the lighter core supports Python 3.9 and up.
• Agents can connect over MCP, a common standard for plugging tools into AI agents.

Keeping history instead of overwriting

The feature our hosts highlighted is bi-temporal storage. In plain terms, the system tracks not only what the agent knew but when it knew it. The project says it keeps changes over time rather than quietly replacing old facts, so an outdated version stays on record once it is superseded.

That matters when a team changes direction. If a note from March says one thing and a note from June says another, you can see both and tell which replaced which. The dashboard has a timeline view with an option to show only superseded versions.

There is a stated limit here. When that view looks at closed history, it favors literal keyword matches and scores at most 500 candidates. The project says matches based only on meaning that fall outside that window can be left out.

The other behavior worth noting is what the project calls grounded recall. Answers cite the memories that support them, or the system abstains when the evidence is weak. For an agent that might otherwise guess confidently, declining to answer is a useful trait.

Jev, Spec Studio and the paid tier

Around the free core sit a few extras. Spec Studio lets you inspect prompt specifications and clusters of related memories, running fixed checks for coverage, ambiguity and conflict. You review its suggestions before anything in memory changes.

Jev is a review helper. In the local dashboard, a "Review with Jev" action checks whether the evidence supports a claim and looks for contradictions, and an MCP tool covers reviews of commands and of whether work is complete. The project states what Jev cannot do: write memories, authorize commands, or certify that a task is finished. Remote calls need your consent each time, plus a label of public or internal for the data.

Local advisory decisions are free for everyone. Pro and Team plans include managed Jev with up to 500 calls per person per day and no personal provider key. The count is per person, not pooled across a team, and failed or interrupted attempts still count. The hosted control plane behind the paid tier is private, while the project itself is under the Apache-2.0 license, which does not grant trademark rights.

What the evidence does and does not show

The project is unusually frank about its benchmarks. The current registered artifact has three deterministic offline fixture runs, which separately look at context size, retrieval quality and grounded decision checks. The authors say these small fixtures do not establish general task performance.

Its size figure is also an estimate of the shape of the data when serialized as JSON. The authors say it is not a measurement of MCP transport or a provider's billing total.

One experimental feature, route selection for recall, works only with your own provider key. On 40 public synthetic tasks it changed no retrieval quality, so the project makes no claim that it improves results.

The takeaway is that Engraphis offers a thoughtful design, not proof that it makes agents measurably better. Treat it as something to test on your own work.