Open source · AI agent memory

The intelligence layer for AI memory.

Genesys is a scoring engine, causal graph, and lifecycle manager that makes AI memory actually work. Open source. Speaks MCP natively.

89.9% on LoCoMo Apache 2.0pip install genesys-memory
Terminal demo: run pip install genesys-memory, then genesys serve --backend obsidian. Genesys uses a local embedder with no API keys and starts an MCP server on http://localhost:8000/mcp with 11 memory tools registered. Calling memory_recall("onboarding preferences") returns 3 scored memories traced to their source.
01 The problem

Why existing memory fails.

Most "memory" is a flat pile of embeddings. At volume, it buries the few things that matter and never forgets the rest.

01

Flat memory doesn't scale

The 500th memory buries the 5 that matter. Vector search returns noise at volume.

02

No forgetting = no intelligence

Without active pruning, your agent drowns in stale context it should have let go of.

03

No causal reasoning

Vector similarity can't answer "why did I choose X?" For that, you need a graph.

02 How it works

Scoring, lifecycle, and a causal graph.

Three mechanisms work together so the right memories surface, the stale ones fade, and every recall can explain itself.

Multiplicative scoring

Earn the recall.

score = relevance
    × connectivity
    × reactivation

Multiplicative, not additive. A zero in any dimension sends the score to zero — no free rides.

Memory lifecycle

Forget on purpose.

ActiveDormantFading
Pruned·Core — never decays

Memories move through states over time. Core memories are promoted and protected; the rest decay unless they keep earning their place.

Causal graph

Answer "why?"

causedcontradictssupportsrefines

Memories connect by typed edges — not just embedding similarity. Traverse the graph to reconstruct how a decision was actually reached.

03 Storage backends

Run it your way.

Same engine, four backends. Start in-memory, scale to Postgres, or keep everything local.

01

In-Memory

Zero dependencies. Try it out in seconds.

02

Postgres + pgvector

Production-ready and scalable.

03

Obsidian vault

Local-first. Your files stay untouched.

04

FalkorDB

Graph-native traversal.

Fully local option: GENESYS_EMBEDDER=local — no API keys needed.
04 Benchmarks

89.9% overall on LoCoMo.

1,540 questions across 10 conversations. Scored against the leading memory systems.

SystemLoCoMo score
Genesys89.9%
SuperLocalMemory87.7%
Zep75.1%
Mem067.1%
05 MCP tools

Every operation is an MCP tool.

Works with Claude Code, Claude Desktop, OpenAI Codex, and any MCP-compliant client.

memory_storeStore a new memory
memory_recallRecall top-scored memories
memory_searchSemantic search over memories
memory_traverseWalk the causal graph
memory_explainExplain why a memory was recalled
memory_statsUsage and lifecycle statistics
pin_memoryPin a memory to prevent decay
unpin_memoryUnpin a memory
delete_memoryPermanently delete a memory
list_core_memoriesList core memories
set_core_preferencesSet core memory categories
06 Quick start

Up and running, fully local.

Obsidian vault setup — no API keys. Four steps from install to a memory-equipped agent.

01 — Install

Add the package

pip install genesys-memory
02 — Configure

Point it at your vault

GENESYS_BACKEND=obsidian OBSIDIAN_VAULT_PATH=/path/to/vault GENESYS_EMBEDDER=local
03 — Run

Start the MCP server

uvicorn genesys.api:app --port 8000
04 — Connect

Wire up your client

{ "mcpServers": { "genesys": { "url": "http://localhost:8000/mcp" } } }

Then tell your agent: always check memory before asking the user to repeat themselves.

07 FAQ

Frequently asked questions.

Short answers to the questions developers ask most about Genesys.

What is Genesys?

Genesys is an open-source AI agent memory engine from Astrix Labs. It combines a multiplicative scoring engine, a causal graph, and a lifecycle manager so the right memories surface, stale ones decay, and every recall can explain itself. It speaks MCP natively and is Apache 2.0 licensed.

How does Genesys score memories?

Genesys scores each memory multiplicatively as relevance multiplied by connectivity multiplied by reactivation. Because the factors are multiplied rather than added, a zero in any single dimension drops the score to zero, so memories must earn recall across every dimension instead of coasting on one strong signal.

Does Genesys require API keys?

No. Genesys can run fully local by setting GENESYS_EMBEDDER=local, which uses an on-device embedder with no API keys required. Paired with the Obsidian vault backend, the entire memory pipeline runs on your own machine while leaving your existing vault files untouched.

How does Genesys compare to Mem0 and Zep?

On the LoCoMo benchmark of 1,540 questions across 10 conversations, Genesys scores 89.9% overall. That leads SuperLocalMemory at 87.7%, Zep at 75.1%, and Mem0 at 67.1%, giving Genesys the highest recall accuracy among the compared memory systems in this evaluation.

Open source

Give your agents memory that actually works.

Apache 2.0 — free and open source · Speaks MCP natively