Agents shouldn't forget everything overnight.

The problem we kept running into

Every new session, a coding agent starts from zero. It re-asks which package manager you use. It re-learns the codebase from scratch. It forgets why the team chose PASETO over JWT six weeks ago, or that one endpoint is deprecated, or which migration is still waiting on review.

Hosted memory services solve part of this, but the fix comes with a catch: your private context leaves your machine and lands on a third party's servers, usually billed per memory stored or per retrieval made. For teams working on regulated, on-prem, or simply private codebases, that trade isn't acceptable.

The Engraphis way

Engraphis takes the opposite approach: a memory layer that runs entirely on your machine, free and Apache-2.0 at the core, with a native MCP server so coding agents plug in directly. No SDK glue, no framework to adopt.

• Local-first and private. Runs offline; the core engine depends only on numpy. No API key required for memory, no per-token cost, no data exfiltration.
• Product WebUI. A dashboard and Memory Inspector to browse, search, and manage your knowledge graph. No other memory engine ships a local product interface like this.
• Agent-native. 18 MCP tools for Claude Code, Cursor, Cline, Zed, and Windsurf.
• Self-maintaining facts. Writes are deterministically conflict-resolved (add, reinforce, or supersede) with no LLM required for the core write path.
• Principled recall. A six-term score over retention, semantic similarity, lexical match, graph proximity, importance, and recency.
• Bi-temporal truth. Contradictions invalidate instead of overwriting. engraphis_why and engraphis_timeline show exactly what was believed, and when.
• Grounded, not guessed. Cited answers or an explicit abstain, with provenance on every memory.
• Code-aware. An AST-powered symbol graph via engraphis_index_repo and engraphis_search_code.
• Sleep-time consolidation. A scheduled job distills recurring episodes and reports its own compaction.

How the memory actually decays

Under the hood, Engraphis models memory the way memory research does: an Ebbinghaus forgetting-curve decay, interaction-aware reinforcement, bi-temporal facts, and hybrid recall that blends vector search, lexical matching, and knowledge-graph proximity. Facts that get reinforced stick around. Facts that go stale fade, the same way a person's memory works, except every step here is inspectable.

Free and open-source at the core

The engine, dashboard, Inspector, MCP server, and governance tools are Apache-2.0 and free, permanently. Paid tiers fund ongoing development: analytics, compliance export, team features. The core recall engine is never crippled to push an upgrade. See the pricing breakdown for what's in each tier.

Where the code lives

Engraphis is developed in the open on GitHub. The core engine sits under engraphis/core, with pluggable backends for embedders, vector indexes, rerankers, and the code graph. The same MemoryService facade backs the dashboard, the Inspector, the MCP server, and the Python library, so what you see in the UI is exactly what your agents get.

"Engraphis" is a trademark of the Engraphis project. The Apache-2.0 license covers the code; it doesn't grant trademark rights.

Get involved

Engraphis is built in public. Read the source, file an issue, or say hello. See the contact page for the fastest way to reach the project.