Ahok Memory

Long-term memory for AI agents

Real memory.
Recall with reasons.

Ahok Memory is a self-hosted memory engine for AI agents. Store once, then retrieve by similarity, recency, and importance — with a trace that shows why every memory surfaced.

Install the engineRead the docs
Runs
Self-hosted
Storage
SQLite or Postgres
Reach
Python, Node, MCP
Recall traceworked example

Query

“What did we decide about the pricing page?”

  1. Pricing v2 approved after the usage-based test.

    0.91
    Episodic
    recency84
    importance62

  2. Brand voice: plain verbs, sentence case.

    0.78
    Semantic
    recency41
    importance88

  3. Standup notes, 3 Nov — pricing copy review.

    0.55
    Episodic
    recency95
    importance30

How a memory comes back

Three steps, in order. The last one is the difference between a memory engine and a vector table.

  1. 1

    Store

    An agent writes a memory with a user id and optional tags. Memories stay partitioned by user, so one person's context never leaks into another's.

  2. 2

    Consolidate

    Episodes become semantic facts. Temporal edges record what happened before what, and reinforcement raises the memories that keep proving useful.

  3. 3

    Recall

    A query is scored by similarity, recency, and importance. The result comes back with the reasons it surfaced, not just a list of nearest vectors.

Install

Use the SDK when you want embedded memory. Run the server when several agents, people, or services need to share it.

  • Local SQLite by default; Postgres for production.
  • Same memory across OpenAI, Anthropic, or a local model.
  • MCP server included for Claude, Cursor, and other clients.

Or run the HTTP server: npx openmemory-js serve

Python

bash
pip install openmemory-py

Node.js

bash
npm install openmemory-js

Then follow the quick start. It takes one file.

Built for inspection

Memory is only useful if you can check it. The data model is visible in the API and in the dashboard.

Sectors
Episodic, semantic, and procedural memories are stored separately, so an event and a learned fact do not compete for the same slot.
Edges
Temporal and semantic links connect memories. Multi-hop questions can follow a chain instead of guessing from a flat index.
Traces
Every recall exposes its score and the factors behind it. When a memory looks wrong, you can see why it was chosen.

Start with one memory. Add the second when you need it.

Open dashboardRead the docs