synap-langgraph

Synap memory integration for LangGraph.

Install

pip install synap-langgraph

Requires langgraph>=1.0, maximem-synap>=0.2.0.

What's in the box

• SynapStore — implements LangGraph's BaseStore for cross-thread long-term memory. Semantic search via store.search(namespace, query=...) routes to sdk.fetch(...), so your graph nodes get Synap-powered recall out of the box.

  • User or customer scope. Pass user_id for private per-user memory, or just a customer_id (no user_id) for a customer-wide shared pool visible to every user in the deployment.
  • All memory types. Reads surface facts and preferences (plus episodes / emotions / temporal events), so stated preferences aren't dropped.
  • Anticipation (optional). Construct with include_conversation_context=True and feed turns via store.record_message(conversation_id, role, content) so just-stated context is in play on the next read. (This lives alongside the BaseStore API — anticipation has no key/value analogue.)

• SynapCheckpointSaver — implements BaseCheckpointSaver with best-effort fuzzy retrieval. Checkpoint writes succeed durably; reads use sdk.fetch which is semantic-search shaped rather than exact KV. Use for observability/audit and demo flows. For production checkpoint durability, pair with SqliteSaver / PostgresSaver.

• create_synap_node — re-exported from synap-langchain for users who discovered our LangGraph support through the LangChain package. This is the canonical home.

• synap_st_prompt — short-term conversation context as a prompt callable for create_react_agent. Prepends Synap's compacted summary + recent turns above your system prompt at every LLM step.

• create_synap_st_node — same short-term context, exposed as a StateGraph node that writes the ST string into state for your LLM node to consume.

Short-term context (compacted conversation, on every LLM step)

LangGraph's built-in memory truncates recent turns to a token budget. Synap's short-term context is the compacted summary + recent turns maintained per conversation by the Synap server — a richer, more token-efficient view of "what happened so far." Drop it into a prebuilt agent or a custom graph; both helpers serve from the SDK's local cache when warm (near-zero overhead) and fall back to the Synap server when cold.

The SDK helper they both wrap is sdk.conversation.context.get_context_for_prompt(conv_id, style=...), which is cache-first whenever the SYNAP_SDK_ST_AUTHORITATIVE flag is on.

A) Prebuilt agent — one-line drop-in

from langgraph.prebuilt import create_react_agent  # or langchain.agents.create_agent
from maximem_synap import MaximemSynapSDK
from synap_langgraph import synap_st_prompt

sdk = MaximemSynapSDK(api_key="sk-...")

agent = create_react_agent(
    model="anthropic:claude-3-5-sonnet-20241022",
    tools=[...],
    prompt=synap_st_prompt(
        sdk,
        conversation_id="conv_abc123",       # required, explicit
        system="You are a helpful agent.",   # your own instructions
        style="narrative",                   # default; also "structured" | "bullet_points"
    ),
)

What the model sees at every step (system message content):

<synap_short_term_context>
... compacted summary + recent turns from Synap ...
</synap_short_term_context>

You are a helpful agent.

B) Custom graph — write ST into state, consume in your own LLM node

from typing import Annotated, TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.graph.message import add_messages
from synap_langgraph import create_synap_st_node

class State(TypedDict):
    messages: Annotated[list, add_messages]
    synap_st: str   # populated by the node

async def my_llm_node(state: State):
    system_text = f"You are a helpful agent.\n\n{state['synap_st']}".strip()
    ...   # build your prompt template with system_text and invoke the model

graph = StateGraph(State)
graph.add_node("st", create_synap_st_node(sdk, conversation_id="conv_abc123"))
graph.add_node("llm", my_llm_node)
graph.add_edge(START, "st")
graph.add_edge("st", "llm")
graph.add_edge("llm", END)

Which to use

Error policy

• SDK failures never crash the graph by default (on_error="fallback"): logged at ERROR via SynapIntegrationError's log path, then the helper degrades to your bare system prompt (or empty state slot).
• Pass on_error="raise" for strict environments that want the failure surfaced as SynapIntegrationError.
• An empty short-term result (no compaction yet and no recent turns) is a legitimate empty case, not a failure — the user's system prompt is preserved as-is.

Conversation ID

Always explicit. We deliberately do not infer it from LangGraph's thread_id because the two namespaces can diverge — your thread might span multiple Synap conversations, or vice versa. For multi-conversation agents, construct one prompt callable per conversation inside your per-run setup.

Quickstart

from langgraph.graph import StateGraph, START, END
from maximem_synap import MaximemSynapSDK
from synap_langgraph import SynapStore, SynapCheckpointSaver

sdk = MaximemSynapSDK(api_key="sk-...")

store = SynapStore(sdk, user_id="alice", customer_id="acme")
saver = SynapCheckpointSaver(sdk, user_id="alice", customer_id="acme")

graph = StateGraph(MyState)
# ... add nodes / edges ...
app = graph.compile(checkpointer=saver, store=store)

# Store usage inside a node:
async def remember(state, runtime):
    await runtime.store.aput(
        ("alice", "preferences"),
        "language",
        {"value": "English"},
    )

Error policy

• Writes (SynapStore.put, SynapCheckpointSaver.put, put_writes) surface SDK failures as SynapIntegrationError. Silent drops would hide ingestion outages.
• Reads (get, search, get_tuple, list) degrade gracefully — they log at ERROR and return None/[] rather than crashing the graph.
• Deletes (SynapStore.delete, SynapCheckpointSaver.delete_thread) warn once and no-op — Synap has no public delete API.

Note on metadata-stripping backends. get/search identify items by custom metadata markers. On instances that atomize content during extraction (e.g. MACA), those markers are stripped, so:

• search falls back to returning the scope-filtered results Synap ranked (semantic retrieval still works). Scope (user_id/customer_id) is enforced at the fetch layer, but sub-namespace isolation within a scope is not. A one-time warning is logged. Set SynapStore(..., semantic_fallback=False) for strict namespace semantics (search returns [] when markers are absent).
• get (exact key) returns None — there's no reliable way to resolve an exact key without the markers. Use search as the read path.

Job/document-level attribution (mapping fragments back to a source id) is not done in the store; build it in app code.

When to use which checkpointer

Cross-thread long-term memory (BaseStore) maps cleanly to Synap — use SynapStore as your default.