synap-microsoft-agent

Synap memory integration for Microsoft Agent Framework (MAF).

Install

pip install synap-microsoft-agent

Requires agent-framework>=1.0, maximem-synap>=0.2.0.

Quickstart

from agent_framework import InMemoryHistoryProvider
from agent_framework.openai import OpenAIChatClient
from maximem_synap import MaximemSynapSDK
from synap_microsoft_agent import SynapContextProvider, SynapHistoryProvider

sdk = MaximemSynapSDK(api_key="sk-...")
client = OpenAIChatClient(model="gpt-4o-mini")

agent = client.as_agent(
    name="MemoryAgent",
    instructions="You are a helpful assistant.",
    context_providers=[
        SynapContextProvider(
            sdk=sdk,
            user_id="alice",
            customer_id="acme",
        ),
        SynapHistoryProvider(
            sdk=sdk,
            user_id="alice",
            customer_id="acme",
        ),
    ],
)

session = agent.create_session()
result = await agent.run("What's my trial expiring?", session=session)

What each provider does

• SynapContextProvider: on every turn, fetches Synap context (facts, preferences, episodes, emotions, temporal events) and appends it as instructions. Records the user + assistant messages back to Synap, and reports the whole turn on the live stream (see below). This is the provider that owns the turn lifecycle. It also adds SynapCompactionMiddleware to every run: once Synap has compacted the conversation, the model sees Synap's compaction and the turns after it instead of the earlier turns your history provider (InMemoryHistoryProvider, Redis, Cosmos...) holds. Until then, or if Synap cannot answer, the messages are untouched; MAF's own compaction stays installed. It works wherever the history provider sits in the list.

• SynapHistoryProvider: keeps the conversation log in Synap. Loads Synap's view on session resume (the compaction, then the turns after it). Subclass of MAF's HistoryProvider, so all its flags (load_messages, store_inputs, store_outputs, store_context_messages) work as documented. Next to SynapContextProvider each turn is still written once, by this provider.

• SynapShortTermContextProvider: injects Synap's compaction of the current conversation into the instructions, refreshed each turn. It adds no recent turns: your history provider already sends them.

You can use either or both; they coexist.

The live stream

The first report opens sdk.instance.listen() if you have not, and this integration reports the whole turn on it (SYNAP_AUTO_LISTEN=false turns that off and the turns go over REST), so Synap's anticipation agent sees more than a question and an answer:

Stream first, REST only as a fallback, never both. The server writes the conversation row from the stream event itself, so reporting a turn and calling record_message would store it twice and extract it twice. Each turn goes out on the stream if one is open and over REST if it is not.

Without a stream (auto-listen off, or it will not open) the user and assistant turns go over REST and the tool, reasoning and result reports are no-ops. No report can raise into your agent run.

A system message is not a turn and is not written: it has no stream event (report_turn would file it as something the assistant said) and record_message takes user or assistant only.

The Agent Harness

MAF's harness (create_harness_agent) is a separate surface with its own memory subsystem, and this package backs both of its storage seams. Needs agent-framework>=1.13; the classes import lazily so the floor above stays at 1.0.

from agent_framework import create_harness_agent
from synap_microsoft_agent import SynapAgentFileStore, create_synap_harness_memory

agent = create_harness_agent(
    client,
    history_provider=create_synap_harness_memory(sdk, user_id="alice", customer_id="acme"),
    file_memory_store=SynapAgentFileStore(sdk, user_id="alice", customer_id="acme"),
)

• SynapMemoryStore — backs the topic notebook (MEMORY.md, topic records, extraction, consolidation). Extraction and consolidation stay MAF's; storage and retrieval become Synap's. MEMORY.md gains a durable recall block that survives restarts and is shared across agents on the same scope.
• SynapAgentFileStore — backs the seven file_memory_* tools. grep searches by regex and by meaning; delete is real.
• create_synap_harness_memory — builds the provider wired correctly. Prefer it.

Two things to know:

1. create_harness_agent takes exactly one history_provider, and both SynapHistoryProvider and the harness memory provider are HistoryProviders. Passing both silently drops one. Use the factory and pass only that.
2. Topic records are held exactly, and by default only for the life of the process. Synap's ingestion rewrites what you submit, and the harness does read-modify-write on records, so reading them back from Synap would corrupt them a little more each turn. Pass a record_store to survive restarts. The content is durable in Synap either way.

Every harness API is @experimental upstream. Pin a tested version and re-run the suite on each MAF minor.

Error policy

• Read-side failures (fetch, get_context_for_prompt) degrade gracefully — logged at ERROR, empty result returned. An outage never crashes an agent turn.
• Write-side failures — SynapContextProvider.after_run logs and swallows (context providers must not raise per MAF's hook contract). SynapHistoryProvider.save_messages surfaces errors as SynapIntegrationError so explicit persistence failures are observable.
• Stream reports never raise and never fall over into your agent run. A report that fails is a report that did not go out, so the turn takes the REST path instead. The one report that is logged at WARNING rather than debug is a refused tool-reporter install, because that costs every tool call and every tool result for the whole run.
• Harness stores — same split. get_index_text degrades to pointer lines with no recall block, because it feeds the system prompt on every turn. write_topic and file_memory_write raise. A topic that is not held raises FileNotFoundError, which is MAF's own not-found contract.

Tests

pytest integrations/synap-microsoft-agent/tests -q

The harness tests skip cleanly when agent-framework<1.13 is installed, so the SDK surfaces stay testable on the floor version. bench/smoke.py drives MAF's real MemoryContextProvider and FileMemoryProvider over both stores — run it with --mock for no network, or with credentials and --scope against a seeded scope.