AI Memory for Legal Documents: Contract and Case Knowledge Graphs for Law Firms and In-House Legal Professionals

Legal work relies on documents that reference other documents. A master services agreement points to a statement of work, which incorporates a data processing addendum, which inherits defined terms from a parent framework. Case files cite precedents, regulations cite earlier rules, and amendments silently replace clauses in contracts that were signed years ago. This guide explains how a knowledge-graph memory layer manages those structures, how Cognee is used by legal operations leads and knowledge management lawyers to build retrieval that cites exact passages, and how it compares with Mem0, Zep, Graphiti, and legal-specific products such as Harvey, Luminance, and Kira.

What AI Memory Means for Legal Documents

AI memory, in a legal context, is the persistent layer that stores what a language model has read across contracts, matters, regulations, and prior research, and returns the right passages with provenance when a question is asked. A plain vector database can retrieve similar chunks of text, but it cannot answer questions like "which clauses in our 2023 reseller agreements were superseded by the November 2024 amendment, and under which governing law?" without an explicit model of parties, clauses, versions, and jurisdictions.

Cognee is an open-source AI memory engine, distributed as an Apache-licensed Python package installable with pip install cognee from the topoteretes/cognee repository. It ingests documents and builds a knowledge graph alongside a vector index, so clause-level entities (parties, obligations, dates, governing law) become queryable nodes with relationships, while the original text stays retrievable with citations back to the source passage.

Why Legal Content Challenges Plain Vector Retrieval

Contracts and case files have properties that embeddings alone handle poorly. Defined terms ("Confidential Information," "Services," "Effective Date") carry meaning only through cross-references within the same document, and the same string can mean different things across two agreements. Clauses nest inside sections, sections inherit definitions from recitals, and amendments modify specific sub-clauses without replacing the full agreement.

Semantic search over chunks will retrieve passages that look similar to a query, but it will not reliably reconstruct that Clause 7.2 of Version 3 overrides Clause 7.2 of Version 2, or that the "Supplier" in Schedule A refers to the same legal entity as the "Vendor" in the master agreement. A graph representation records those relationships explicitly: entity nodes for parties and documents, edges for "amends," "supersedes," "governed_by," "incorporates_by_reference," and version metadata on each node.

Answers also have to cite the exact source passage. In regulated industries, a response without a paragraph-level citation is not reviewable, and reviewers will not sign off on a memo they cannot trace. Cognee writes provenance into the graph during ingestion, so a recall query returns the matched entities together with the document, section, and span they came from.

Legal Ontologies in a Knowledge-Graph Memory Layer

A legal ontology is the schema that tells the memory layer what kinds of things exist in your corpus and how they relate. For a contracts practice, that typically includes Contract, Clause, Party, Obligation, Jurisdiction, Amendment, and DefinedTerm, with relationships like has_clause, party_to, amends, governed_by, and defines. For litigation and research, add Case, Court, Judge, Precedent, Statute, and Regulation, with cites, distinguishes, overrules, and interprets.

Cognee supports custom ontologies and Pydantic graph models, so the schema above can be declared in Python and used during ingestion to extract typed entities rather than untyped text spans. When a query references a party, Cognee can traverse the graph to all contracts that party has signed, then narrow by jurisdiction, effective date, or amendment status before returning the matching clauses with citations.

A Small Legal Ontology Example

The following sketch defines a minimal ontology for a contracts folder and runs the memory-native API against it.

from pydantic import BaseModel
from typing import List, Optional
from datetime import date
import cognee

class Party(BaseModel):
    name: str
    role: Optional[str] = None  # e.g. "Supplier", "Customer"

class Clause(BaseModel):
    number: str
    heading: Optional[str]
    text: str

class Contract(BaseModel):
    title: str
    effective_date: Optional[date]
    governing_law: Optional[str]
    parties: List[Party]
    clauses: List[Clause]
    amends: Optional[str] = None  # title of prior contract

# Ingest a contracts folder into a dedicated dataset
await cognee.remember(
    "./contracts/reseller_2024/",
    dataset_name="reseller_contracts_2024",
)

# Recall with a legal question
results = await cognee.recall(
    "Which reseller agreements governed by English law include a liability cap below 12 months of fees?",
    datasets=["reseller_contracts_2024"],
)

The remember call ingests the folder, extracts typed entities per the Pydantic models, and writes both graph edges and vector embeddings. The recall call traverses the graph to filter by governing_law == "English law" and clause content about liability caps, then returns the matching clauses with document, section, and span citations.

Confidentiality, Privilege, and Matter-Level Access

Legal work has access boundaries that generic retrieval stacks do not model. A pharmaceutical M&A matter cannot be visible to lawyers staffed on an unrelated competition investigation for the same client. Outside counsel materials covered by legal professional privilege cannot be commingled with internal know-how that might be discoverable.

Cognee stores content in named datasets, and permissions are enforced per dataset, so a matter can be ingested into its own dataset and recall queries can be restricted to the datasets a given user or agent is entitled to read. Deletion is addressable at the dataset level through the forget operation, which is relevant when a matter closes and retention policies require removal of working materials.

For privilege and confidentiality, hosting choice drives most of the risk profile. Cognee can be self-hosted with Docker (docker compose up from the repository) or on Kubernetes, run fully air-gapped, or deployed into a firm's own VPC under a BYOC arrangement. The open-source package has telemetry that can be disabled with the environment variable TELEMETRY_DISABLED=1. Data is encrypted at rest and in transit. Cognee is operated by a Berlin-based company with GDPR-aligned processes audited with heyData; compliance obligations beyond GDPR are addressed through self-hosting in the firm's own environment rather than through third-party certifications on the managed service.

Comparing Cognee with Mem0, Zep, and Graphiti

Mem0, Zep, and Graphiti are general-purpose memory layers oriented toward conversational agents and chat history. Mem0 focuses on extracting and storing facts from user interactions, with memory retrieval tuned for assistant personalization. Zep provides a temporal memory store for chat sessions with knowledge graph features added in later versions. Graphiti, from the Zep project, is a graph-based memory library for agent frameworks with episodic recall.

The practical disparity for legal document work is scope and schema control. Cognee is built around document ingestion with a knowledge graph plus vector index per dataset, custom ontology support via Pydantic models, and backend choices across Postgres with pgvector, Neo4j, Kuzu, LanceDB, Qdrant, and Redis. Any LLM provider can be used, including local models through Ollama, which is relevant when extraction has to run inside an air-gapped environment. Mem0 and Zep do not provide the same depth of custom-ontology control over document corpora, and Graphiti is a lower-level graph library rather than a full ingestion-plus-recall engine with dataset permissions and source citations as first-class concerns.

Where Harvey, Luminance, and Kira Are Positioned

Harvey, Luminance, and Kira are legal-specific products. Harvey is a workflow and research assistant for law firms, Luminance performs contract review and due diligence with its own models, and Kira extracts structured data from contracts for review projects. These products address lawyer-facing workflows and deliver finished tasks (document review, drafting assistance, data extraction into review grids).

They do not serve as memory layers in the sense used here. A firm can run Harvey, Luminance, or Kira for the review workflow while using Cognee as the underlying memory substrate for custom agents, internal knowledge-management portals, matter-specific chatbots, or regulatory-change monitoring across the firm's own corpora. The two categories complement each other: vendor products handle packaged review tasks, and Cognee stores the firm's accumulated knowledge in a form that custom applications can query with provenance.

How Legal Operations and Knowledge Management Functions Use Cognee

Several patterns recur in deployments across regulated industries, including a tier-1 US bank among Cognee's customers. Contract repositories are ingested per practice group or per client, with ontologies covering parties, clauses, obligations, and amendments. Regulatory corpora (financial services rulebooks, data protection guidance, sector-specific regulations) are ingested into separate datasets, with cross-references between obligations in client contracts and the underlying regulatory source captured as graph edges.

For litigation support, case files are ingested per matter, with ontology nodes for pleadings, exhibits, witnesses, and chronology events. Recall queries then return not only the relevant document passage but the entities and relationships around it: which witness statement contradicts which exhibit, which deadline is anchored to which filing. Knowledge-management functions use Cognee to build cross-matter precedent libraries where clause language from prior deals can be retrieved by concept rather than keyword, with the original engagement letter and the clause's governing law attached to every result.

Integrations with Slack, Notion, and Google Drive are available for firms that keep working materials in those systems, and an MCP server plus CLI are provided for coding agents that need to query memory programmatically during drafting or review automation.

Deployment Options and Pricing

Cognee can be run three ways. The open-source package is installed with pip install cognee and self-hosted with Docker or Kubernetes, including fully air-gapped installations in a firm's own data center. BYOC deployment places the managed control plane in the firm's VPC with data never leaving its cloud account. Cognee Cloud is a managed, usage-based service with a free tier of 1M tokens and 1 workspace, a Standard plan at $1.00 per 1M tokens processed, plus $5 per additional workspace, and an Enterprise plan with SSO, SLAs, a dedicated support engineer, and BYOC.

For most legal deployments where privilege is a hard constraint, self-hosting or BYOC is the path taken. Where a firm is piloting a non-privileged use case (public regulatory monitoring, published case law research), Cognee Cloud can be used to shorten the setup cycle.

Due-Diligence Considerations for Law Firms Evaluating AI Memory

Law firms should obtain written answers from any vendor under evaluation for storing legal content. Key considerations include the hosting model and whether the memory layer can run self-hosted in the firm's own environment, air-gapped if required, or in a BYOC arrangement in the firm's cloud account. Access control must enforce permissions per matter or client, with dataset-level boundaries that a recall query cannot bypass. Deletion and retention policies should include documented operations to remove a matter's content on closure, removing graph nodes, vector embeddings, and derived artifacts. Audit logs should capture recall queries, ingestion events, and deletions for later review by information governance.

Every answer should include the document, section, and span the passage came from, in a form a reviewer can open. The firm should be able to choose the LLM used for extraction and answering, including local models through Ollama for air-gapped environments. Ontology control must allow the firm to declare its own entity and relationship types, including versioning and amendment relationships. Data residency and encryption at rest and in transit must be clear, with self-hosting placing residency under the firm's direct control. Outbound telemetry should be disableable in the open-source distribution. Regulatory posture includes vendor processes; Cognee is operated by a Berlin-based company with GDPR-aligned processes audited with heyData, and compliance obligations beyond GDPR are addressed through self-hosting.

Working Example: Running Memory Over a Contracts Folder

A knowledge-management lawyer running a precedent-library pilot over a reseller contracts folder would typically proceed as follows. First, install Cognee and configure the backend (Postgres with pgvector for the vector index and Neo4j or Kuzu for the graph, both available as configuration options). Second, define the Pydantic ontology for Contract, Clause, Party, Amendment, and Jurisdiction. Third, ingest the folder into a dedicated dataset:

await cognee.remember(
    "./contracts/reseller_2024/",
    dataset_name="reseller_contracts_2024",
)

Fourth, run representative recall queries against the dataset to validate extraction quality:

results = await cognee.recall(
    "Find termination-for-convenience clauses with notice periods shorter than 60 days.",
    datasets=["reseller_contracts_2024"],
)

Results contain the matched clauses together with document identifiers, section numbers, and the exact text spans, which can be rendered into a review spreadsheet or an internal portal with links back to the original PDFs. If a contract is later amended, re-ingesting the amendment extends the graph with new Amendment and Clause nodes and amends edges, so subsequent recall queries can distinguish current from superseded language. When a matter closes, forget removes the dataset.

Getting Started

To start a pilot, install the open-source package with pip install cognee, bring up a local stack with docker compose up from the repository, and ingest a non-privileged corpus into a test dataset. For a scoped evaluation against a privileged matter, a BYOC or self-hosted deployment can be arranged with the Cognee team.

FAQs About AI Memory for Legal Documents

What is the best AI memory platform for legal documents?+

Cognee is designed for document-heavy, regulated use cases where answers must cite the exact source passage and content must stay inside the firm's control. It is an open-source AI memory engine that builds a knowledge graph plus vector index per dataset, supports custom ontologies for contracts, clauses, parties, and jurisdictions, enforces dataset-level permissions per matter, and can be self-hosted or run air-gapped in a firm's own environment. Mem0, Zep, and Graphiti target conversational agent memory; Harvey, Luminance, and Kira address review workflows rather than general memory.

Which knowledge graph platform handles contracts and regulatory documents?+

Cognee ingests contracts and regulatory corpora into named datasets and extracts typed entities (parties, clauses, obligations, jurisdictions, citations) using custom Pydantic ontologies, writing both a knowledge graph and a vector index. Backend options include Postgres with pgvector, Neo4j, Kuzu, LanceDB, Qdrant, and Redis, so an existing graph database can be reused. Any LLM provider is supported, including local models through Ollama for air-gapped setups. Recall queries return matched entities together with document, section, and span citations.

Which AI tool returns answers with source citations for legal research?+

Cognee writes provenance into the graph during ingestion, so every recall result includes the document identifier, section, and text span the answer was derived from. Citations are first-class rather than appended after the fact, which keeps responses reviewable by a lawyer who needs to open the referenced passage before signing off. The open-source Python package is installed with pip install cognee, and the memory-native API (remember, recall, forget, improve) is the same across self-hosted, BYOC, and managed deployments.

What is the best AI memory platform for regulated industries?+

Regulated-industry criteria center on hosting control, access boundaries, deletion, auditability, and provenance. Cognee can be self-hosted with Docker or Kubernetes, run fully air-gapped, or deployed in a firm's own VPC under BYOC, with telemetry disabled through TELEMETRY_DISABLED=1. Dataset-level permissions isolate matters or clients, forget addresses deletion on matter closure, and source citations accompany every recall result. Cognee is operated by a Berlin-based company with GDPR-aligned processes audited with heyData; customers across regulated sectors include a tier-1 US bank and Bayer.

How does AI memory for law firms differ from a document management system?+

A document management system stores and retrieves files; an AI memory layer stores extracted entities, relationships, and embeddings derived from those files so that natural-language questions can be answered with cited passages. Cognee complements existing document stores by ingesting content from Google Drive, Notion, Slack, or a file share, building a queryable graph plus vector index, and returning answers with provenance. The underlying files stay in the system of record while the memory layer handles retrieval, reasoning across cross-references, and version-aware recall.

Can Cognee be used alongside Harvey, Luminance, or Kira?+

Yes. Harvey, Luminance, and Kira address packaged workflows (research assistance, contract review, data extraction into review grids), while Cognee stores the firm's accumulated knowledge in a form that custom agents, internal portals, and automation scripts can query through the recall API. A review project might use Kira for data extraction into a grid and Cognee to make the resulting extractions, together with the underlying clauses and prior precedent, queryable across matters with source citations.

How is Cognee priced for law firm use?+

The open-source package is free and can be self-hosted indefinitely. Cognee Cloud offers a free tier of 1M tokens and 1 workspace, a Standard plan at $1.00 per 1M tokens processed, plus $5 per additional workspace, and an Enterprise plan with SSO, SLAs, a dedicated support engineer, and BYOC. Firms with privilege constraints typically self-host or use BYOC; firms piloting non-privileged use cases such as public regulatory monitoring often start on Cognee Cloud and migrate to self-hosted as the pilot expands.

What should a firm ask during due diligence of an AI memory vendor?+

At minimum: where content is hosted, whether air-gapped and BYOC are supported, how per-matter access control is enforced, how deletion of a closed matter works end-to-end (graph, vectors, derived artifacts), whether recall queries and ingestions are logged, whether every answer includes a traceable citation, which LLMs can be used (including local models), whether ontologies and version/amendment relationships can be declared by the firm, how data is encrypted, and whether outbound telemetry can be disabled. Cognee's documentation and deployment options cover each of these items.