Relevant context gets lost
Naive history replay grows without bounds. Stale or superseded facts can keep influencing new decisions.
OPEN-SOURCE DEVELOPER INFRASTRUCTURE
Memory for agents.
Boundaries for runs.
Give Python agents context that can be traced, tool calls that can be bounded, and failures that are easier to understand.
Synapse brings persistent memory, provenance-aware retrieval, opt-in execution controls, and privacy-minimized run traces into one self-hostable layer.
Python SDK · HTTP API · LangChain · LlamaIndex
RUN / SYN-8F2A
Limit reached
customer_lookup
Started just now · 4 events
STEPS04 / 04
TOOL CALLS02 / 03
STATUSStopped
Run startedPolicy attached · metadata only
Memory retrieval3 evidence records · 2 packed
Tool call failedcustomer_lookup · timeout
Repeated call blockedConfigured failure ceiling reached
No raw prompt or tool arguments shownEVENTS 04
Context with provenanceSource · time · trust · packing
Explicit run limitsSteps · calls · deadline
BUILT FOR DEVELOPERS WHO SHIP AGENTS
PYTHONHTTP APIPOSTGRES + PGVECTORAPACHE 2.0 SOURCE
Release status: PyPI package 1.3.1 is published. Repository source targets 1.4.0; that version is prepared but not yet published to PyPI.
Understand the difference
THE PROBLEM
Agents can fail in ways
Agents can fail in ways
a better prompt cannot fix.
Useful agents need more than a longer context window. They need a way to manage what they remember, constrain what they do, and inspect what happened.
Tool calls can repeat
Retries and loops can multiply side effects. A model saying “done” does not prove an external action succeeded.
Failure evidence is fragmented
Without a connected run record, it is harder to understand which context was selected and where execution stopped.
WHAT SYNAPSE BRINGS TOGETHER
One layer. Four useful primitives.
Start with the part your application needs. Keep your existing agent framework and deployment decisions.
MEMORY
Memory that has a lifecycle
Ingest, deduplicate, retrieve, update, supersede, and retract information instead of replaying an ever-growing transcript.
- Hybrid lexical + vector retrieval with rank fusion
- Temporal validity and source/provenance metadata
- Token-budgeted context packing
INGEST→RETRIEVE→PACK→USE
EXECUTION
Opt-in run boundaries
Route selected work through SynapseRun to enforce step, tool-call, and cooperative deadline limits, with repeated-failure controls.
Explore run controls →EVIDENCE
Traces tied to context
Connect run events with retrieval IDs, memory evidence, ranking signals, token costs, and packing decisions—without storing raw prompts or tool payloads.
Explore trace model →DEPLOYMENT & CONTROL
Keep the infrastructure in your environment
Use the Python SDK or HTTP API. Develop locally with SQLite; use PostgreSQL and pgvector for production-shaped deployments. Tenant boundaries, scoped keys, and operational requirements are documented—not hidden behind a hosted-only workflow.
HOW IT FITS
Keep your agent.
Keep your agent.
Add the missing layer.
Synapse is designed to sit beside your existing application—not replace your orchestration framework.
YOUR APPLICATION
✳
Python agent
Tools · model calls · business logic
LangChainLlamaIndexCustom
SDK / HTTP
SYNAPSE LAYER
S
Memory + run evidence
Memory & retrievalRun boundariesTrace events
Storage
YOUR DATA LAYER
SQLite / PostgreSQL
pgvector · tenant isolation
Operator-managed infrastructureConceptual architecture. Exact deployment topology depends on your environment and configuration.
FIRST STEP
Try the memory layer
Try the memory layer
on your machine.
The published package supports a local SQLite evaluation path. It is the easiest way to test the write-and-retrieve flow before provisioning production infrastructure.
Read the developer quickstartPython 3.10+ · local-only evaluation · no external database required
TERMINAL
python -m pip install synapse-memory-engine==1.3.1 PYTHON
from synapse_memory import SynapseClient
with SynapseClient(
"http://127.0.0.1:8008",
tenant_id="local-dev",
) as client:
saved = client.create_memory(
"The app uses SQLite for local evaluation.",
category="fact",
)
result = client.retrieve("Which database is used locally?")
print(saved.id)
print(result.fused_context)✓ Example follows the published SDK quickstart
TRUST BY BEING PRECISE
Security details
Security details
should be explicit.
Synapse documents its boundaries so developers can decide whether its data model and deployment requirements fit their workload.
Read the security policy ↗Field-level encryption has a defined scope
The authoritative content field is Fernet-encrypted. The plaintext search projection, text-search derivative, and embeddings remain readable by design; protect database volumes and backups accordingly.
Tenant isolation depends on deployment roles
PostgreSQL RLS must be paired with the documented non-owner, non-superuser runtime role and migrate-then-run setup.
Run evidence is metadata-only
Trace events are designed to exclude prompts, raw tool arguments, tool results, and memory text. Applications remain responsible for safe event metadata.
Validation status is not overstated
CI and synthetic fault-injection tests are not customer-workload proof, a security certification, or a production-readiness guarantee.
CURRENT PROJECT STATUS
Built in the open.
Built in the open.
Validated in stages.
We want developers to see what exists, what is still being validated, and where their feedback can change the direction.
Published Python package1.3.1
Repository source target1.4.0 prepared
Independent security reviewPending
Real-world pilot validationOpen
The 1.4.0 source additions are not part of the published 1.3.1 package. See the version notes.