gligabao

Minimal memory system for AI. A small, self-hosted MCP/HTTP server that turns text into a temporal knowledge graph and answers questions from it.

Long-term memory for your AI application

gligabao is a small, self-hosted MCP/HTTP server: use it over its HTTP API, or as an MCP server for Claude, Claude Code and other MCP clients. You send it free-form text; it turns that text into a knowledge graph. Later you ask questions in natural language and it answers from what it stored, citing the facts it used.

Quickstart

Build the binary, start the server and remember your first fact in minutes.

How it works

Remember, query and the bi-temporal graph underneath.

API reference

Every endpoint: remember, query, datasets, graph export, ontologies, timeseries.

Embed in Go

Use gligabao as a library inside your own Go program.

The two core calls

Remember

curl -X POST localhost:7878/v1/gb/remember \
  -H 'content-type: application/json' \
  -d '{"user_id": "u1", "text": "Alice works at Acme and lives in Paris."}'

1. Remember

   An LLM extracts entities and edges (relations such as works_at) from the text. Entities are deduplicated, and each fact is anchored to the artifact it came from and records when it was true.

2. Query

   The query finds relevant entities and facts, by keywords and by the similarity of their embeddings; the surrounding subgraph of live facts is expanded by one hop, the chunks of long texts that are close to the question are added as passages, and an LLM answers using only that context. The response includes the evidence behind the answer.

3. Store

   SQLite (a single file, the default) or PostgreSQL, plus an embedded or hosted vector store.

Explore

Temporal facts

Bi-temporal edges keep history instead of deleting it.

Datasets

Isolated, per-user containers for memory.

Sessions

Trace every fact back to the run it came from.

Images

Read text and themes from images.

Timeseries

Anchor values over time to entities and facts.

Geo

Coordinates for the entities that are places.

MCP

remember and query as tools for MCP clients.

Evals

End-to-end checks recorded as ATIF trajectories.