AutoGen and AG2¶

"AutoGen" is now three things. This page records what each is and what the adapters implement.

Why AG2 gets tools and not a KnowledgeStore: that protocol requires atomic append returning a byte offset, read_range over unbounded files, and directory listing. HMS stores at most 64 KiB per document, has no atomic append and no listing endpoint, so an implementation could not honor the contract (write-ahead logs would exceed one document within a few turns). AG2 agents get semantic memory through tools instead, which is what a vector store is for.

Install¶

pip install "holographic-integrations[autogen]"   # autogen-agentchat 0.4+ Memory
pip install "holographic-integrations[ag2]"       # AG2 1.x tools

The two extras conflict (different frameworks); install one per environment.

AutoGen quickstart¶

from holographic_integrations.autogen import HolographicMemory
from holographic_integrations.core import HmsConfig

memory = HolographicMemory("assistant_notes", hms=HmsConfig(url="http://127.0.0.1:8080", tenant_id="agents", input_dim=384), k=5)
agent = AssistantAgent("helper", model_client=client, memory=[memory])
await memory.add(MemoryContent(content="Prefers dark roast", mime_type=MemoryMimeType.TEXT, metadata={"topic": "food"}))

update_context queries with the latest message and adds a system message in the same format as ListMemory. query(q, k=, score_threshold=, filter=) takes a MongoDB-style filter over the MemoryContent.metadata you stored. Results carry metadata["id"] and metadata["score"] (cosine). name is the collection, so several agents can share one tenant. clear() deletes the collection's records.

Content types: text, Markdown and JSON dict (round-trips as a dict). bytes and images raise ValueError. Supply metadata["id"] to make an add idempotent.

Serialization: dump_component() / load_component() work (provider = holographic_integrations.autogen.HolographicMemory). The config holds name, URL, tenant, dimension, k, threshold, overfetch and max_fetch. The API key (read from HMS_API_KEY on load) and a custom embedder are never serialized.

AG2 quickstart¶

from ag2 import Agent
from holographic_integrations.ag2 import HolographicMemoryToolkit

agent = Agent("assistant", config=cfg, tools=[HolographicMemoryToolkit(hms_config, k=5)])

Tools: remember(text, metadata?) -> id, recall(query, k?, filter?) -> [{id, text, metadata, score}], forget(id) -> bool (omit with allow_forget=False). The toolkit's .backend is the AsyncHmsBackend for direct use.

Knobs¶

Both take an HmsConfig: url, api_key, tenant_id, collection, embedder (default offline HashingEmbedder; set input_dim to match --input-dim), text_queries, zero_trust_key, overfetch, max_fetch.

Filter mapping¶

Limits: k is at most max_fetch (default 100, the server's --max-top-k); no listing endpoint, so there is no "all memories" read; text and metadata are 64 KiB each.

Tested¶

• autogen-agentchat / autogen-core / autogen-ext 0.7.5, Python 3.12: Memory ABC, add/query round trips, filters, JSON content, collections isolated in one tenant, update_context into a real BufferedChatCompletionContext, a real AssistantAgent(memory=[...]) run with ReplayChatCompletionClient asserting the memory text reached the model call and a MemoryQueryEvent was emitted, component dump/load (tests/autogen/test_autogen_memory.py).
• ag2 1.1.2, Python 3.12: a real Agent with scripted TestConfig calling remember then recall, asserting the tool result reached the model and the record is in the service; forget; the toolkit survives AG2's deep copy of tools (tests/autogen/test_ag2_toolkit.py).
• Examples: examples/autogen_memory.py, examples/ag2_memory.py (offline, run against a live hms-server).
• Not run: real LLM providers; Microsoft Agent Framework.