OMEM: trustworthy memory for AI agents

OMEM is a memory layer for AI agents that tracks beliefs over time and handles contradictions instead of silently overwriting them. This is the official Python SDK. It has no third-party dependencies, so it installs instantly and won't clash with anything else in your environment.

Install

pip install omem-infrastructure

The import is short:

from omem import Memory

Run a server in one command

Installing also gives you an omem-server command that starts the full OMEM server (engine and API) locally, with no extra setup:

omem-server

It runs on http://127.0.0.1:8787 and stores its data in an omem-data folder in your current directory. Use a different port with omem-server 9000.

Quickstart

from omem import Memory

mem = Memory(api_key="omem_sk_...", project="proj_...")

# Remember a grounded fact. The agent and entity are created on first use.
mem.remember(agent="support-agent", about="customer:123",
             claim="prefers_annual_billing")

# Ask what's believed. The engine resolves contradictions for you.
print(mem.believes(about="customer:123", claim="prefers_annual_billing"))
# -> BELIEVED_TRUE

# Recall everything known about an entity, with provenance.
for m in mem.recall(about="customer:123")["memories"]:
    print(m["proposition"], m["state"])

# See why something is believed, with the full provenance chain.
mem.why("a_...")

Cross-agent memory

Memory is private to an agent by default, and you decide what to share.

# Private to one agent. Only agent-a can recall it.
mem.remember(agent="agent-a", about="acme", claim="secret_deal=1",
             scope="agent:agent-a")

# Shared across the whole project. Every agent can recall it.
mem.remember(agent="agent-a", about="acme", claim="tier=enterprise", scope="org")

# Shared with a named team.
mem.remember(agent="agent-a", about="acme", claim="ae=jane", scope="team:sales")

# Promote an existing memory to a wider scope later.
mem.share(assertion_id="a_...", scope="org")

Use it as an MCP server

Installing also gives you an omem-mcp command that speaks MCP over stdio and exposes three safe tools: omem_recall, omem_observe, and omem_why.

OMEM_API_KEY=omem_sk_... OMEM_BASE_URL=https://... OMEM_AGENT=support-agent omem-mcp

Point your MCP client (such as Claude Desktop) at that command. The agent identity is fixed at the process level, so a model can't reach into another agent's private memory.

Self-healing

# Report a failure and let OMEM's policy-gated recovery loop handle it.
mem.healing.report(component="db-pool", error_type="ECONNRESET")
mem.healing.handle(error={"component": "db-pool", "error_type": "ECONNRESET"})
mem.healing.health()   # aggregated component health

How it fits together

Every method maps onto one operation or query in the OMEM engine. The SDK adds authentication, retries on 5xx errors, typed errors (OmemError.reason_code exposes codes like R_DANGLING), automatic registration of agents and entities, and cross-agent scope control. It doesn't invent any new memory behavior of its own; the engine remains the single source of truth.