ilma

Framework-agnostic agent memory system.

Postgres + pgvector backend. MCP-native. Hermes Agent, Claude, Cursor, Codex — any MCP client.

pip install ilma-agent
ilma init
ilma status

What ilma is

ilma stores what your agents know — and makes it retrievable across sessions, frameworks, and machines.

• 8 memory surfaces: memories, wiki, journal, skills, metrics, kanban, observability, sessions

• Hybrid retrieval: vector + FTS + chunk-level reranking

• Graph layer (Apache AGE): cross-entity traversal over Memory / Wiki / Skill vertices with SHARES_TAG, CO_OCCURS, REFERENCES_WIKI, and USES_SKILL edges. Rebuild on demand; expand recall hits via expand_graph=True on ilma_recall.

• MCP server: ilma-mcp — works with any MCP client

• HTTP API: REST endpoints behind your own reverse proxy

• CLI: ilma init, ilma search, ilma remember, ilma graph rebuild, ilma doctor

• Postgres + pgvector: proven, backup-friendly, multi-client

Related MCP server: mesh-memory

Quick start

# Install
pip install ilma-agent

# Start the ilma-db container (Postgres 18 + pgvector + pg_cron + timescaledb + age)
# See the "Production deployment" section below for the full recipe with
# named volumes, restart policy, and resource limits.
docker run -d --name ilma-db --restart always \
    --cpus=2 --memory=4g \
    -e POSTGRES_DB=ilma \
    -e POSTGRES_USER=ilma \
    -e POSTGRES_PASSWORD=change-me \
    -v ilma-pg-data:/var/lib/postgresql/data \
    -v ilma-pg-init:/docker-entrypoint-initdb.d \
    -p 127.0.0.1:5432:5432 \
    ghcr.io/brotal-llc/ilma-pg:latest

# (Optional) Start an embedder — Ollama with bge-m3 (1024-dim) pre-pulled
docker run -d --name ilma-ollama --restart always \
    --cpus=4 --memory=8g \
    -v ilma-ollama-data:/root/.ollama \
    -p 127.0.0.1:11434:11434 \
    ghcr.io/brotal-llc/ilma-ollama:latest
# (the image's entrypoint pulls `bge-m3` on first start if missing)

# Initialize (9-step wizard: Postgres, extensions, schemas, embedder)
ilma init

# Store a memory
ilma remember "User prefers dark mode" --tags user,preference

# Search
ilma search "dark mode preference"

# Check health
ilma doctor

Production deployment

The canonical recipe: two containers (ilma-db for Postgres, ilma-ollama for embeddings) plus the ilma-agent Python package running as the API/MCP service. All three names are stable so systemd / docker-compose / k8s can reference them by name.

ilma-db (Postgres 18 + extensions)

docker run -d --name ilma-db --restart always \
    --cpus=2 --memory=4g \
    --health-cmd="pg_isready -U $$POSTGRES_USER -d $$POSTGRES_DB" \
    --health-interval=10s --health-timeout=5s --health-retries=5 \
    -e POSTGRES_DB=ilma \
    -e POSTGRES_USER=ilma \
    -e POSTGRES_PASSWORD=change-me \
    -v ilma-pg-data:/var/lib/postgresql/data \
    -v ilma-pg-init:/docker-entrypoint-initdb.d \
    -p 127.0.0.1:5432:5432 \
    ghcr.io/brotal-llc/ilma-pg:latest

• --name ilma-db — stable name. The ilma API and any tooling reference it by name.

• --restart always — survives host reboots, Docker daemon restarts.

• --cpus=2 --memory=4g — Postgres 18 + 5 extensions needs at least 1 CPU and 2GB to be comfortable; bump these up for large wiki/journal tables.

• Named volumes — ilma-pg-data holds the database cluster (survives container removal); ilma-pg-init exposes /docker-entrypoint-initdb.d so you can drop custom .sh / .sql files into it without rebuilding the image.

• -p 127.0.0.1:5432:5432 — bind to localhost only. The ilma API runs on the same host and connects via host.docker.internal or the unix socket. If you need cross-host access, terminate Postgres at a firewall or use a WireGuard tunnel rather than exposing 5432 publicly.

The image ships with pgvector, pg_trgm, ltree, timescaledb, pg_cron, and apache-age preinstalled. The ilma Python package's initialize_schema() creates the ilma schema and tables on first API connection — no SQL migration step needed.

ilma-ollama (embedder, optional but recommended)

ilma defaults to bge-m3 (1024-dim) as the embedding model. The image's entrypoint pulls bge-m3 on every container start — idempotent when the named volume /root/.ollama already has the model on disk.

docker run -d --name ilma-ollama --restart always \
    --cpus=4 --memory=8g \
    --health-cmd="curl -fsS http://localhost:11434/api/tags || exit 1" \
    --health-interval=15s --health-timeout=5s --health-retries=10 \
    --health-start-period=180s \
    -v ilma-ollama-data:/root/.ollama \
    -p 127.0.0.1:11434:11434 \
    ghcr.io/brotal-llc/ilma-ollama:latest

The first boot takes 1-3 minutes while the model downloads (~2.2GB for bge-m3). The healthcheck waits 180s before starting to account for that. After the first run, subsequent restarts are instant because the model is in the named volume.

Why not bake the model into the image? BuildKit's sandbox can't reliably reach a long-running background ollama serve from inside a single RUN step, so we can't pre-pull at build time. Pulling on every start (with named-volume caching) is faster on subsequent runs.

Image size: ~6.6GB. The ollama/ollama base image is 6.6GB on disk regardless of variant (the -rocm tags are larger). There's no slimmer official ollama image — the binary + Go runtime + Ubuntu base is just that big. If size matters, consider:

• A custom fork that strips ROCm/CUDA from the official binary (~1.5GB reduction, but you'd need to maintain it).

• A pure-Python replacement using llama.cpp Python bindings (~500MB, but no HTTP API, no model registry).

• For now we accept the 6.6GB floor — ollama is feature-complete and well-supported.

ilma-agent (the API/MCP service)

Run the API/MCP service however you run Python services today (systemd, docker, supervisor). All it needs is ILMA_DSN and optionally an embedder URL:

# .env or systemd Environment= entries
ILMA_DSN=postgresql://ilma:change-me@ilma-db:5432/ilma
ILMA_EMBED_PROVIDER=ollama
ILMA_EMBED_BASE_URL=http://ilma-ollama:11434
ILMA_EMBED_MODEL=bge-m3
ILMA_EMBED_DIM=1024

See ilma init for the interactive wizard that writes a complete .env.

docker-compose equivalent

If you prefer compose, here's the same recipe as a single file (typically lives in ~/infra/ilma/compose.yaml on the deploy host, NOT in this repo):

services:
  ilma-db:
    image: ghcr.io/brotal-llc/ilma-pg:latest
    container_name: ilma-db
    restart: always
    cpus: 2
    memory: 4G
    environment:
      POSTGRES_DB: ilma
      POSTGRES_USER: ilma
      POSTGRES_PASSWORD: change-me
    volumes:
      - ilma-pg-data:/var/lib/postgresql/data
      - ilma-pg-init:/docker-entrypoint-initdb.d
    ports:
      - "127.0.0.1:5432:5432"
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U $$POSTGRES_USER -d $$POSTGRES_DB"]
      interval: 10s
      timeout: 5s
      retries: 5

  ilma-ollama:
    image: ghcr.io/brotal-llc/ilma-ollama:latest
    container_name: ilma-ollama
    restart: always
    cpus: 4
    memory: 8G
    volumes:
      - ilma-ollama-data:/root/.ollama
    ports:
      - "127.0.0.1:11434:11434"
    healthcheck:
      test: ["CMD", "curl", "-fsS", "http://localhost:11434/api/tags"]
      interval: 15s
      timeout: 5s
      retries: 10
      start_period: 180s

volumes:
  ilma-pg-data:
  ilma-pg-init:
  ilma-ollama-data:

Architecture

MCP client (Claude, Cursor, Hermes, etc.)
    ↓  stdio / HTTP
ilma MCP server / HTTP API
    ↓  psycopg3
Postgres + pgvector (+ pg_cron + timescaledb + apache-age)
    ├─ ilma.memories
    ├─ ilma.wiki
    ├─ ilma.journal
    ├─ ilma.skills
    ├─ ilma.metrics
    ├─ ilma.kanban
    ├─ ilma.observability
    ├─ ilma.sessions
    └─ ag_catalog.ilma_graph (derived view — Memory / Wiki / Skill vertices + edges)

The ilma_graph lives in Apache AGE. It's a derived view over the relational state in ilma.* and is rebuilt on demand via ilma graph rebuild. See src/ilma/core/graph.py and the ilma-age-graph skill for design notes.

Docker images

The ilma-pg image ships with: pgvector, pg_trgm, ltree, timescaledb, pg_cron, and apache-age. The ilma Python package's initialize_schema() creates the ilma schema and tables on first API connection — no SQL migration step needed.

The ilma-ollama image is a thin wrapper over ollama/ollama that runs ollama pull bge-m3 on first start. Use it as the ILMA_EMBED_BASE_URL target for any ilma installation that wants self-hosted embeddings.

The legacy ghcr.io/skb50bd/hermes-memory/hermes-postgres image is deprecated. New deployments should use ghcr.io/brotal-llc/ilma-pg.

Project layout

.
├── src/ilma/                  # Python package
│   ├── core/                  # pure logic (memory, wiki, retrieval, graph)
│   ├── storage/               # PostgreSQL backend (pgvector, AGE)
│   ├── api/                   # CLI + MCP + HTTP API
│   ├── adapters/              # hermes-memory provider shim
│   └── plugins/               # hermes-agent plugin entry points
├── pg/                        # Postgres image (Dockerfile + entrypoint scripts)
│   ├── Dockerfile             # pgvector + pg_cron + timescaledb + age
│   └── bin/                   # ilma-pg-entrypoint.sh, ilma-pg-init.sh, ilma-pg-cron.sh
├── ollama/                    # Ollama embedder image (Dockerfile + entrypoint)
│   ├── Dockerfile             # ollama/ollama + bge-m3 (1024-dim) auto-pulled
│   └── bin/                   # ilma-ollama-entrypoint.sh
├── tests/                     # unit + integration
│   ├── unit/                  # 226 tests, no DB
│   └── integration/           # 12 tests via Testcontainers (ilma-pg:latest)
├── Dockerfile                 # ilma CLI image (the API service)
├── pyproject.toml
└── .github/workflows/ilma.yml # single gated CI: lint, test, security, tag,
                              # build (ilma CLI), build-pg (Postgres),
                              # build-ollama (Ollama+bge-m3), release

No docker/ or infra/ directories. Local-dev Docker Compose and Caddy configs live in ~/infra/ on the deployment host, not in this repo. The CI workflow publishes ready-to-pull images. See the "Production deployment" section above for the canonical recipe (container names ilma-db / ilma-ollama, named volumes, restart policy, resource limits).

MCP tools (31)

The ilma MCP server exposes 31 tools, including the graph surface:

• ilma_recall(query, ..., expand_graph=False, graph_hops=1) — graph-aware recall.

• ilma_wiki_search(query, ..., expand_graph=False, graph_hops=1) — graph-aware wiki search.

• ilma_graph_rebuild(min_shared_tags=2) — drop and rebuild the AGE graph.

• ilma_traverse(kind, src_id, max_hops=2, edge_types=None, limit=50) — bounded BFS.

• ...plus 27 others (memory, wiki, journal, skills, metrics, observability, sessions, kanban, audit, doctor, migrate, repair).

Development

Install dev dependencies and set up pre-commit:

make install
pip install pre-commit
pre-commit install

Run all checks locally:

make all

Run only the integration tests against the ilma-pg image:

uv run pytest tests/integration/ -v --tb=short

(Requires Docker — Testcontainers spins up ghcr.io/brotal-llc/ilma-pg:latest automatically.)

Project status

In active development. See PLAN.md for the execution roadmap.

License

MIT