anura-graph

Python client for the Anura Memory API.

Anura Memory provides two memory products for AI agents:

• GraphRag — Knowledge graph with automatic triple extraction, deduplication, and hybrid retrieval
• FilesRag — Markdown file storage with heading-based chunking and semantic search

Installation

pip install anura-graph

Quick Start

from graphmem import GraphMem

mem = GraphMem(api_key="gm_your_key_here")

# --- GraphRag ---
mem.remember("Alice is VP of Engineering at Acme Corp")
ctx = mem.get_context("What does Alice do?")

# --- FilesRag ---
mem.write_file("/notes/standup.md", "# Standup\n## 2026-02-21\n- Shipped auth module")
results = mem.search_files("auth module")

Configuration

from graphmem import GraphMem, RetryConfig

mem = GraphMem(
    api_key="gm_your_key_here",
    base_url="https://anuramemory.com",  # default
    retry=RetryConfig(
        max_retries=3,       # default
        base_delay=0.5,      # seconds, default
        max_delay=10.0,      # seconds, default
        retry_on=[429, 500, 502, 503, 504],  # default
    ),
    timeout=30.0,  # seconds, default
)

The client can be used as a context manager:

with GraphMem(api_key="gm_your_key_here") as mem:
    mem.remember("Alice works at Acme")

API Reference

GraphRag

remember(text) -> RememberResult

Extract knowledge from text and store as triples in the graph.

result = mem.remember("Einstein was born in Ulm, Germany")
print(result.extracted_count)  # 1
print(result.merged_count)     # 1

If a new fact contradicts an existing one on a single-valued predicate (e.g., LIVES_IN, WORKS_AT), the old value is replaced and the conflict is returned:

r = mem.remember("Alice now works at Microsoft")
# r.conflicts → [ConflictResolution(subject_name="alice", predicate="WORKS_AT",
#                 old_object="google", new_object="microsoft", resolution="auto_recency")]

get_context(query, options?) -> ContextResult

Retrieve context from the knowledge graph.

from graphmem import ContextOptions

# JSON format (default)
ctx = mem.get_context("Einstein")
print(ctx.entities)  # [{ "name": "Albert Einstein" }, ...]

# Markdown format (ideal for LLM system prompts)
ctx = mem.get_context("Einstein", ContextOptions(format="markdown"))
print(ctx.content)  # "- Albert Einstein -> BORN_IN -> Ulm..."

# Hybrid mode (graph + vector + communities)
ctx = mem.get_context("Einstein", ContextOptions(mode="hybrid"))

search(entity, type_filter=None) -> SearchResult

Find an entity and its direct (1-hop) connections. In fuzzy mode, filter results by entity type.

result = mem.search("Alice")
print(result.edges)

# Search with type filter
result = mem.search("alice", type_filter="person")

ingest_triples(triples) -> IngestResult

Ingest pre-formatted triples directly (no LLM extraction).

from graphmem import Triple

result = mem.ingest_triples([
    Triple(subject="TypeScript", predicate="CREATED_BY", object="Microsoft"),
])

get_graph(type_filter=None) -> GraphData

Get the full graph (nodes, edges, communities). Optionally filter by entity type.

# Filter graph by entity type
people = mem.get_graph(type_filter="person")
orgs = mem.get_graph(type_filter="person,organization")

delete_edge(id, blacklist=False)

Delete an edge. Optionally blacklist to prevent re-creation.

update_edge_weight(id, weight=None, increment=None)

Set or increment an edge's weight.

delete_node(id)

Delete a node and all its connected edges.

export_graph() -> ExportData

Export the graph as portable JSON.

import_graph(data) -> ImportResult

Import a graph export (merges, does not delete existing data).

list_communities() -> list[Community]

List all detected communities.

detect_communities() -> DetectCommunitiesResult

Run Louvain community detection + LLM summarization.

FilesRag

write_file(path, content, name=None) -> WriteFileResult

Create or update a markdown memory file. Files are automatically chunked by ## headings and indexed for semantic search.

result = mem.write_file(
    "/docs/architecture.md",
    "# Architecture\n\n## Backend\nNode.js with Prisma...\n\n## Frontend\nNext.js...",
)
print(result.file.id)       # "clxx..."
print(result.chunk_count)   # 3
print(result.created)       # True

If a file already exists at the given path, its content is replaced and re-indexed.

list_files() -> list[MemoryFile]

List all files in the current project.

files = mem.list_files()
for f in files:
    print(f"{f.path} ({f.size} bytes)")

read_file(id) -> FileWithContent

Get a file with its full content.

file = mem.read_file("file_id")
print(file.content)  # full markdown

update_file(id, content, name=None) -> WriteFileResult

Update a file's content (re-chunks and re-indexes).

result = mem.update_file("file_id", "# Updated content\n...")

delete_file(id)

Delete a file and all its indexed chunks.

search_files(query, limit=None, file_id=None) -> list[FileSearchResult]

Semantic search across file chunks. Pass file_id to scope the search to a single file.

results = mem.search_files("authentication flow", limit=5)
# Search within a single file:
# results = mem.search_files("auth", file_id="file_abc123")
for r in results:
    print(r.file["path"])
    for chunk in r.chunks:
        print(f"  {chunk.heading_title} ({chunk.score:.2f})")
        print(f"  {chunk.excerpt}")

Projects

Traces

Blacklist

Conflict Log

conflicts = mem.list_conflicts(limit=10)
for c in conflicts:
    print(f"{c.subject_name} {c.predicate}: {c.old_object} → {c.new_object}")

Pending Facts

Usage

get_usage() -> UsageInfo

usage = mem.get_usage()
print(usage.tier)                  # "FREE"
print(usage.current_facts)         # 42
print(usage.current_file_count)    # 3
print(usage.current_file_storage)  # 1024

Health

health() -> HealthResult

Error Handling

from graphmem import GraphMem, GraphMemError

try:
    mem.read_file("nonexistent")
except GraphMemError as e:
    print(f"API error {e.status}: {e}")
    print(e.body)  # raw response body

Rate Limiting

After each request, rate limit info is available:

mem.remember("some fact")
print(mem.rate_limit.remaining)  # requests remaining
print(mem.rate_limit.limit)      # total allowed per window
print(mem.rate_limit.reset)      # unix timestamp when window resets

Types

All types are exported from the top-level package:

from graphmem import (
    # GraphRag
    RememberResult, ConflictResolution, ConflictLogEntry,
    ContextResult, SearchResult, Triple,
    GraphNode, GraphEdge, GraphData, Community,
    # FilesRag
    MemoryFile, FileWithContent, FileSearchResult, WriteFileResult,
    # Config
    RetryConfig, RateLimitInfo, UsageInfo,
)

License

MIT