stigmem-plugin-cognee-adapter

Bridges Stigmem with topoteretes/Cognee — a knowledge-graph / memory layer that ingests documents and conversations, extracts entities and relationships via LLM, and exposes them as a queryable graph.

This package is experimental and opt-in. Installing it makes the cognee-adapter plugin discoverable through the stigmem.plugins entry-point group; host applications still choose when to call the adapter.

Federation model

stigmem holds structured atomic facts across agents: typed key-value triples (entity, relation, value) with provenance and scope. Cognee builds a rich relational knowledge graph from unstructured input, making facts accessible via semantic search and graph traversal.

The two stores complement each other:

This adapter exposes two seams:

1. assert_to_cognee(fact, dataset) — when stigmem asserts a fact, push it into Cognee as structured text. Cognee cognifies the text into a typed graph node/edge, making the relationship discoverable through semantic queries that cross fact boundaries.

2. query_from_cognee(scope, query) — run a Cognee semantic search and return results as stigmem-compatible fact-shaped records, ready to merge or display alongside native stigmem facts.

Design

The adapter is intentionally thin:

• It serialises each fact to a structured text string (entity:X | relation:Y | value:V | …) that Cognee can parse back into graph triples.
• It applies Cognee config (set_llm_config, set_vector_db_config) lazily from environment variables on first use — no global side effects at import time.
• Synchronous wrappers (assert_to_cognee, query_from_cognee) call asyncio.run() so callers without an event loop can use the adapter directly. The _async variants are available for callers that already run an event loop.
• batch_assert_to_cognee defers cognify until after all facts are staged — preferred for bulk ingestion because each cognify call invokes the LLM extraction pipeline.

Files

Installation

python -m pip install 'stigmem-plugin-cognee-adapter>=0.1.0,<2.0.0'

The adapter no longer publishes a cognee extra because Cognee currently pulls diskcache==5.6.3, which is affected by CVE-2025-69872 unsafe pickle deserialization. Install Cognee separately only in deployments where the operator has accepted and mitigated that upstream cache-directory risk.

Requirements

• Python ≥ 3.11
• stigmem-py: pip install stigmem-py (or from workspace)
• cognee: optional live runtime dependency installed by the operator; unit tests and plugin discovery do not require it.
• An LLM backend accessible by Cognee (default: OpenAI — set OPENAI_API_KEY or use COGNEE_LLM_PROVIDER + COGNEE_LLM_API_KEY for other providers)

Environment variables

# stigmem node
STIGMEM_URL=http://localhost:8765
STIGMEM_API_KEY=sk-your-key            # optional
STIGMEM_SOURCE_ENTITY=agent:my-agent  # entity URI for assertions

# Cognee — adapter config
COGNEE_STIGMEM_DATASET=stigmem         # dataset name (default: stigmem)

# Cognee — LLM backend (applied lazily on first assert/query)
COGNEE_LLM_PROVIDER=openai
COGNEE_LLM_MODEL=gpt-4o-mini
COGNEE_LLM_API_KEY=sk-your-openai-key  # or set OPENAI_API_KEY directly

# Cognee — vector DB (default: local LanceDB)
COGNEE_VECTOR_DB_PROVIDER=lancedb
COGNEE_VECTOR_DB_PATH=.cognee_db

Usage

Push a single fact into Cognee

from stigmem_plugin_cognee.adapter import StigmemCogneeAdapter

bridge = StigmemCogneeAdapter.from_env()

# fact dict mirrors FactRecord.model_dump()
fact = {
    "entity": "user:alice",
    "relation": "memory:role",
    "value": {"type": "string", "v": "engineer"},
    "source": "agent:my-agent",
    "scope": "company",
    "confidence": 1.0,
}

bridge.assert_to_cognee(fact, dataset="stigmem")

Batch insert (preferred for multiple facts)

bridge.batch_assert_to_cognee(list_of_fact_dicts, dataset="stigmem")
# cognify is called once after all facts are staged

Query the Cognee graph

results = bridge.query_from_cognee(
    scope="company",
    query="What role does alice have?",
)
for rec in results:
    print(rec["entity"], rec["relation"], rec["value"])

Async usage

import asyncio

async def main() -> None:
    bridge = StigmemCogneeAdapter.from_env()
    await bridge.assert_to_cognee_async(fact, dataset="stigmem")
    results = await bridge.query_from_cognee_async("company", "Who owns project:loom?")
    print(results)

asyncio.run(main())

End-to-end demo

cd stigmem/experimental/cognee-adapter
STIGMEM_URL=http://localhost:8765 \
  COGNEE_LLM_PROVIDER=openai \
  COGNEE_LLM_MODEL=gpt-4o-mini \
  COGNEE_LLM_API_KEY=sk-... \
  python demo.py

The demo asserts six related facts about the Loom project into stigmem, pushes them into Cognee's graph, then queries the graph with natural-language questions to verify the relationships are discoverable.

Enable

The adapter has no node-global behavior gate at v0.1.0. Enable it in the host application by installing the package and importing stigmem_plugin_cognee.adapter.StigmemCogneeAdapter.

python -m pip install 'stigmem-plugin-cognee-adapter>=0.1.0,<2.0.0'
stigmem plugins list

Disable

Remove the adapter from the host application path and restart the process that loads plugins. If it was installed only for this integration, uninstall it:

python -m pip uninstall stigmem-plugin-cognee-adapter

Test

cd experimental/cognee-adapter
python -m pytest tests/ -v

No live stigmem node, Cognee instance, or LLM API key required — the cognee module is fully mocked in the test suite.

Uninstall

python -m pip uninstall stigmem-plugin-cognee-adapter

Search types

Cognee exposes several search variants. Pass search_type to query_from_cognee:

INSIGHTS is recommended for structured fact retrieval; GRAPH_COMPLETION for narrative answers that synthesise across multiple facts.

Invariants

• Idempotency: re-asserting the same fact text is safe — Cognee deduplicates at the graph level. stigmem facts are immutable; each new assert gets a new id, so re-pushing the same logical fact creates a duplicate graph node. Applications that need update semantics should query stigmem for the latest fact before pushing.
• Blast radius: Cognee is a secondary enrichment layer. Failures in assert_to_cognee or query_from_cognee do not affect stigmem availability; callers should catch exceptions and degrade gracefully.
• Partition behaviour: if Cognee is unavailable, stigmem continues to serve exact facts. The adapter does not retry internally — callers own the retry / circuit-breaker policy.