hermes-mongodb-memory 0.1.0

MongoDB memory provider plugin for Hermes Agent — hybrid search, time-decay, TTL forgetting, and entity graph traversal in a single backend.

MongoDB Memory Provider for Hermes Agent

Hybrid search, time-decay relevance, TTL-driven forgetting, and entity graph traversal — served from a single MongoDB cluster (self-hosted or Atlas free tier).

Why MongoDB?

Existing Hermes memory providers each pick one storage angle: local SQLite (Holographic), cloud-only API (Mem0, Supermemory, RetainDB), knowledge graph (Hindsight), or vector-only stores. MongoDB collapses all of them into one backend.

Capability	This plugin	Closest alternative
Native hybrid (vector + BM25) ranking	✅ Atlas $rankFusion	RetainDB ($20/mo cloud-only)
Self-hosted and cloud with one URI swap	✅	OpenViking (AGPL-3.0)
TTL-driven automatic forgetting	✅ Mongo TTL indexes	None
Entity graph traversal without a graph DB	✅ $graphLookup	Hindsight (separate KG)
Time-decay relevance built into search	✅ in-pipeline $exp	None
Free tier with cloud durability	✅ Atlas free tier	Holographic (local-only)
Cross-device sync	✅ via Atlas	Cloud providers only

In other words: vector search, BM25, structured queries, TTL, and graph traversal are all the same database — no glue between SQLite and a vector store, no separate graph engine.

Requirements

• Python ≥ 3.10
• A MongoDB 7.0+ cluster, either:
  • Atlas free tier (M0) — recommended; enables $rankFusion hybrid search and managed vector indexes
  • Self-hosted (community / enterprise) — works with the portable BM25 + cosine fallback path
• (Optional) an embedding provider key — OpenAI or Voyage AI. Without one, the plugin runs as a tuned BM25 + entity-graph store.

Install

pip install hermes-mongodb-memory
# Optional vector backends:
pip install "hermes-mongodb-memory[openai]"
pip install "hermes-mongodb-memory[voyage]"
pip install "hermes-mongodb-memory[all]"   # both

Setup

Path 1 — Guided (recommended)

hermes memory setup        # pick "mongodb" from the provider list
hermes mongodb init-indexes

The wizard prompts for connection URI, database, embedding provider, and tenant scope, then writes:

• $HERMES_HOME/mongodb.json — non-secret values
• $HERMES_HOME/.env — HERMES_MONGODB_URI and any embedding API key

Path 2 — Manual

hermes config set memory.provider mongodb
echo 'HERMES_MONGODB_URI="mongodb+srv://USER:[e-mail]/?retryWrites=true"' >> ~/.hermes/.env
hermes mongodb init-indexes

Path 3 — Local development

git clone https://github.com/alexbevi/mongodb_hermes_memory.git
cd mongodb_hermes_memory
pip install -e ".[dev,all]"
# Hermes auto-discovers plugins in ~/.hermes/plugins/<name>/
ln -s "$(pwd)/hermes_mongodb_memory" ~/.hermes/plugins/mongodb
docker compose up -d                            # local Mongo 7 replica set
export HERMES_MONGODB_URI="mongodb://localhost:27017/?replicaSet=rs0"

Atlas free tier (zero-cost path)

1. Create an Atlas account and a free M0 cluster.
2. Network Access → Add your IP (or 0.0.0.0/0 for development).
3. Database Access → Create a user.
4. Connect → Drivers → copy the mongodb+srv://... string.
5. Paste it into hermes memory setup (or set HERMES_MONGODB_URI).
6. (Optional, enables hybrid search) Create the search indexes — see Atlas Search indexes below.

Configuration

$HERMES_HOME/mongodb.json — env vars override file values.

Key	Default	Description
connection_uri	—	MongoDB URI (env: HERMES_MONGODB_URI)
database	hermes_memory	Database name
embedding_provider	none	none / openai / voyage
embedding_api_key	—	Provider API key (env: HERMES_MONGODB_EMBEDDING_API_KEY)
embedding_model	text-embedding-3-small	Embedding model name
embedding_dim	1536	Must match your Atlas Vector Search index
tenant_scope	per-workspace	global / per-workspace / per-profile
auto_extract	false	Run regex extractor at session end
extraction_llm	false	Use Hermes' configured LLM for richer extraction
time_decay_half_life_days	30	Half-life for relevance decay (0 disables)
default_ttl_days	0	Default TTL for new memories (0 = never)
hybrid_vector_weight	0.6	Vector vs BM25 weight (0–1)
prefetch_limit	5	Memories injected into each turn
turn_ttl_days	30	TTL for raw turn capture (0 = forever)

Tools

The model can call six tools while the provider is active:

Tool	Purpose
mongo_remember	Store a memory with category, entities, tags, optional TTL
mongo_search	Hybrid query (vector + BM25 + time-decay) — call before answering personal questions
mongo_recall	Walk the entity co-occurrence graph for related memories
mongo_forget	Remove or set short TTL on a memory by id
mongo_profile	Summary of latest preferences, projects, decisions
mongo_reflect	Cluster recent memories by entity for cross-category synthesis

Indexes

hermes mongodb init-indexes creates the following on memories:

Name	Shape	Why
tenant_category_recent	{tenant_id, category, created_at}	Fast mongo_profile and per-category list scans
tenant_entities	{tenant_id, entities}	Entity recall + $graphLookup start point
tenant_tags (sparse)	{tenant_id, tags}	Tag filtering without bloating non-tagged docs
ttl_expires_at (sparse, TTL=0)	{expires_at}	MongoDB TTL monitor evicts ephemeral memories
content_text	{content: "text"}	Portable BM25 fallback for self-hosted clusters

Plus on turns:

• tenant_session_turn — {tenant_id, session_id, turn_idx}
• turns_ttl — TTL on expires_at

Atlas Search indexes

For the hybrid-search fast path, create these in the Atlas UI (or via the Atlas Admin API):

// Atlas Search index "hermes_memory_text"
{ "mappings": { "fields": { "content": { "type": "string" } } } }

// Atlas Vector Search index "hermes_memory_vector"
{
  "fields": [
    { "type": "vector", "path": "embedding", "numDimensions": 1536, "similarity": "cosine" },
    { "type": "filter", "path": "tenant_id" },
    { "type": "filter", "path": "category" },
    { "type": "filter", "path": "entities" }
  ]
}

Without these, the plugin transparently falls back to portable BM25 + in-process cosine.

Testing

# Unit tests (mongomock-backed, no docker required)
pytest tests/ -v --cov=hermes_mongodb_memory

# Integration tests against a real MongoDB
docker compose up -d
export MONGODB_TEST_URI="mongodb://localhost:27017"
pytest tests/test_integration.py -v -m integration

# Lint
ruff check .

Architecture

flowchart TB
    subgraph Hermes["Hermes Agent (host)"]
        Register["register(ctx)"]
        Provider["MongoDBMemoryProvider"]
        Prefetch["prefetch / queue_prefetch"]
        SyncTurn["sync_turn"]
        Tools["tools (6)"]
        SessionEnd["on_session_end"]

        Searcher["HybridSearcher"]
        Embedder["EmbeddingClient<br/>(openai / voyage / none)"]
        Store["MongoStore"]
        Dispatcher["ToolDispatcher"]
        Regex["RegexExtractor"]
        LLM["LLMExtractor (optional)"]

        Register --> Provider
        Provider --> Prefetch
        Provider --> SyncTurn
        Provider --> Tools
        Provider --> SessionEnd

        Prefetch --> Searcher
        Searcher --> Embedder
        SyncTurn --> Store
        Tools --> Dispatcher
        Dispatcher --> Store
        Dispatcher --> Searcher
        SessionEnd --> Regex
        SessionEnd --> LLM
        Regex --> Store
        LLM --> Store
    end

    subgraph Mongo["MongoDB / Atlas"]
        Memories[("memories")]
        Turns[("turns")]
        Profiles[("profiles")]
        Entities[("entities")]

        RankFusion["$rankFusion<br/>$vectorSearch + $search (BM25)"]
        GraphLookup["$graphLookup<br/>(entity traversal)"]
        TTL["TTL indexes<br/>(automatic forgetting)"]
    end

    Searcher -.->|aggregation| RankFusion
    Store -.->|CRUD + indexes| Memories
    Store -.-> Turns
    Store -.-> Profiles
    Store -.-> Entities
    Dispatcher -.->|mongo_recall| GraphLookup
    Memories -.- TTL

    classDef host fill:#f5f5f5,stroke:#444,color:#222;
    classDef store fill:#e6f4ea,stroke:#2e7d32,color:#1b5e20;
    classDef op fill:#fff3e0,stroke:#ef6c00,color:#5d4037;
    class Register,Provider,Prefetch,SyncTurn,Tools,SessionEnd,Searcher,Embedder,Store,Dispatcher,Regex,LLM host;
    class Memories,Turns,Profiles,Entities store;
    class RankFusion,GraphLookup,TTL op;

Background daemon threads handle prefetch and sync_turn so a flaky cluster never blocks the conversation. A 5-failure / 120-second circuit breaker pauses Mongo writes when something is wrong upstream.

Releasing

Publishing is scripted in scripts/publish.sh:

scripts/publish.sh build     # build sdist + wheel only (dry run)
scripts/publish.sh test      # build, lint, test, upload to TestPyPI
scripts/publish.sh prod      # same, upload to PyPI, tag v<version>

The script enforces a clean working tree, asserts that pyproject.toml and hermes_mongodb_memory/__init__.py agree on the version, runs ruff + pytest, calls python -m build, and uses twine to upload. Configure credentials in ~/.pypirc or via TWINE_USERNAME=__token__ / TWINE_PASSWORD=pypi-....

Contributing

PRs welcome. Conventional Commits required (feat(scope): ..., fix(scope): ...); tests required for new behavior; coverage gate is 80%.

License

MIT — same as Hermes Agent.